6f:T5ac, Yeah, that's just yeah, those are the leads in and then the main one the dedicated circuit It's still hung up there worked up there. Yeah, so the main circuit still good. It's right there So there's still the main circuit left, but they Cut the fiber leading to me cut the fiber leads for all the other parts. Look at this. This is as a travesty Alright, so this is where the demarcation starts. This is where we have actually our first sets of UPS batteries So, and notice this is all off. There we go, there's all the UPS batteries and things like that. Once again, I love Dennis. Now walking through, there's offices down there, so this building does have offices. Here's a little maintenance server room. There's a little, what do we have here? Old foundry switch mounted. It's so weird to be in a data center like this when it's just quiet. Now, data centers are not built with a normal floor so this may look like a floor. Can we move these Ali? Yes. Give you a hint what's going on down here. There we go. Here's how the raised floors work in these. So this right here, not much light so you can, but you can get an idea. This is so all the cable goes underneath and then we're going to go over to each of these and show you how they come through. So we're going to open this up and then you have, this is how the pull throughs come each at the bottom kind of going through the racks. And here's an idea of how easy these panels just kind of...70:T5ac, Yeah, that's just yeah, those are the leads in and then the main one the dedicated circuit It's still hung up there worked up there. Yeah, so the main circuit still good. It's right there So there's still the main circuit left, but they Cut the fiber leading to me cut the fiber leads for all the other parts. Look at this. This is as a travesty Alright, so this is where the demarcation starts. This is where we have actually our first sets of UPS batteries So, and notice this is all off. There we go, there's all the UPS batteries and things like that. Once again, I love Dennis. Now walking through, there's offices down there, so this building does have offices. Here's a little maintenance server room. There's a little, what do we have here? Old foundry switch mounted. It's so weird to be in a data center like this when it's just quiet. Now, data centers are not built with a normal floor so this may look like a floor. Can we move these Ali? Yes. Give you a hint what's going on down here. There we go. Here's how the raised floors work in these. So this right here, not much light so you can, but you can get an idea. This is so all the cable goes underneath and then we're going to go over to each of these and show you how they come through. So we're going to open this up and then you have, this is how the pull throughs come each at the bottom kind of going through the racks. And here's an idea of how easy these panels just kind of...71:T5ac, Yeah, that's just yeah, those are the leads in and then the main one the dedicated circuit It's still hung up there worked up there. Yeah, so the main circuit still good. It's right there So there's still the main circuit left, but they Cut the fiber leading to me cut the fiber leads for all the other parts. Look at this. This is as a travesty Alright, so this is where the demarcation starts. This is where we have actually our first sets of UPS batteries So, and notice this is all off. There we go, there's all the UPS batteries and things like that. Once again, I love Dennis. Now walking through, there's offices down there, so this building does have offices. Here's a little maintenance server room. There's a little, what do we have here? Old foundry switch mounted. It's so weird to be in a data center like this when it's just quiet. Now, data centers are not built with a normal floor so this may look like a floor. Can we move these Ali? Yes. Give you a hint what's going on down here. There we go. Here's how the raised floors work in these. So this right here, not much light so you can, but you can get an idea. This is so all the cable goes underneath and then we're going to go over to each of these and show you how they come through. So we're going to open this up and then you have, this is how the pull throughs come each at the bottom kind of going through the racks. And here's an idea of how easy these panels just kind of...72:T5ac, Yeah, that's just yeah, those are the leads in and then the main one the dedicated circuit It's still hung up there worked up there. Yeah, so the main circuit still good. It's right there So there's still the main circuit left, but they Cut the fiber leading to me cut the fiber leads for all the other parts. Look at this. This is as a travesty Alright, so this is where the demarcation starts. This is where we have actually our first sets of UPS batteries So, and notice this is all off. There we go, there's all the UPS batteries and things like that. Once again, I love Dennis. Now walking through, there's offices down there, so this building does have offices. Here's a little maintenance server room. There's a little, what do we have here? Old foundry switch mounted. It's so weird to be in a data center like this when it's just quiet. Now, data centers are not built with a normal floor so this may look like a floor. Can we move these Ali? Yes. Give you a hint what's going on down here. There we go. Here's how the raised floors work in these. So this right here, not much light so you can, but you can get an idea. This is so all the cable goes underneath and then we're going to go over to each of these and show you how they come through. So we're going to open this up and then you have, this is how the pull throughs come each at the bottom kind of going through the racks. And here's an idea of how easy these panels just kind of...73:T5ac, Yeah, that's just yeah, those are the leads in and then the main one the dedicated circuit It's still hung up there worked up there. Yeah, so the main circuit still good. It's right there So there's still the main circuit left, but they Cut the fiber leading to me cut the fiber leads for all the other parts. Look at this. This is as a travesty Alright, so this is where the demarcation starts. This is where we have actually our first sets of UPS batteries So, and notice this is all off. There we go, there's all the UPS batteries and things like that. Once again, I love Dennis. Now walking through, there's offices down there, so this building does have offices. Here's a little maintenance server room. There's a little, what do we have here? Old foundry switch mounted. It's so weird to be in a data center like this when it's just quiet. Now, data centers are not built with a normal floor so this may look like a floor. Can we move these Ali? Yes. Give you a hint what's going on down here. There we go. Here's how the raised floors work in these. So this right here, not much light so you can, but you can get an idea. This is so all the cable goes underneath and then we're going to go over to each of these and show you how they come through. So we're going to open this up and then you have, this is how the pull throughs come each at the bottom kind of going through the racks. And here's an idea of how easy these panels just kind of...74:Ta32,This is an **NVIDIA H100** data-center accelerator; in the Dedicated offering, the corresponding SKU is **H100-80GB NVL**, intended for high-throughput workloads. [[SRC:0006]] **What we run on H100** - Use H100 as the primary tier for **large-sequence / long-context inference**, including document summarization and retrieval-augmented generation, where its memory capacity and matrix performance benefit large KV caches and long sequences. [[SRC:0003]] - It is also the preferred tier for **latency-critical endpoints**. A100 is used as a fallback/overflow tier; in one EU burst design, the canary was 4 H100s plus 6 preemptible A100s. [[SRC:0007]] - For a mixed dedicated deployment, the proposed allocation was **50% H100** for long-context work, **40% L4** for short-context chat/embeddings, and **10% A100** for legacy or heavy-KV-cache compatibility and burst overflow. [[SRC:0003]] **Dedicated pools and regions** - The Dedicated H100-80GB NVL launch covers **eu-west-3 (Paris)** and **ap-south-1 (Mumbai)**, with reservations available for committed pools. [[SRC:0006]] - A separate APAC dedicated standby/canary pool is defined in **asia-southeast1**, using one `h100-80gb` node; capacity there is described as very limited. [[SRC:0008]] - An **eu-central1** burst pool uses H100 as the primary latency tier with A100 fallback capacity. [[SRC:0007]] - Broader capacity planning prioritizes **NA—us-east and us-west—for low-latency Dedicated pools**, followed by EU for compliance/isolation; APAC is staged while H100 performance is validated. [[SRC:0001]] - Region economics favor reserved **us-west1** H100 capacity as the baseline; reserved europe-west3 is indexed at 118 versus 100 in us-west1, while asia-southeast1 on-demand is indexed at 203, with egress also noted. [[SRC:0009]] **Compared with A100** - H100 provides higher HBM capacity and matrix performance for long sequences and large-batch KV cache workloads; A100 is the more balanced compatibility/overflow option. [[SRC:0003]] - In one internal Llama2-13B, 2,048-context benchmark, H100 delivered about **2× A100 throughput** at comparable batch sizes and similar latency. [[SRC:0006]] - In another conservative sizing estimate, H100 was approximately **1,200 vs. 850 tokens/sec** for short sequences and **160 vs. 110 tokens/sec** for long generation, relative to A100—showing that the gain depends on workload and serving configuration. [[SRC:0004]] - The planning estimate prices H100 at **$7.20/GPU-hour** versus **$3.20/GPU-hour** for A100, so H100 is the performance tier rather than the lowest-cost tier. [[SRC:0001]]75:Td43,eng-infra Sanaa (infra PM): Quick heads-up — we're opening two new Dedicated capacity SKUs next week: H100-80GB NVL for high-throughput workloads and A10G-mid for lower-cost inference. Regions coming online: eu-west-3 (Paris) and ap-south-1 (Mumbai). Reservations open today for committed pools: https://redwood.internal/dedicated/reserve :rocket: Javier (capacity eng): Nice. Do we have rough perf/latency numbers vs existing A100 pools? Sanaa (infra PM): Short answer: H100 ~2x throughput vs A100 for Llama2-13B at 2048 context in our internal harness. Latency similar on equivalent batch sizes. We'll publish the full perf matrix on the console by EOD. Mei (sales): Can existing Dedicated customers swap capacity into these SKUs or is it new add-on only? Customers asking about migration path. Sanaa (infra PM): It's opt-in. No forced migrations. Customers can request add-on pools or renew with H100-backed capacity. Support will help with any migration steps. Priya (support): I'll draft a KB + customer email template. Link to draft when ready. Luca (sre): Any maintenance window to bring these nodes up? Need to plan oncall. Sanaa (infra PM): Minimal control-plane churn. We'll do a canary in us-west-2 this Friday 2024-04-26 02:00 UTC (short restarts for a subset of agents), then broad rollout Mon-Tue 2024-04-29/30. Expected downtimes: two ~5m rolling restarts (control plane agents), app plane unaffected. deploy-bot will post exact timing. deploy-bot: Scheduled rollout created: Canary us-west-2 -> Broad rollouts 2024-04-29 to 2024-04-30. Release ref: go/release/1234. Rajesh (platform): Heads-up for quota owners: new quota buckets and SKU codes will appear in the portal. If teams need increases for demos or GA customers, please request via quotas channel or the portal. Javier (capacity eng): SDK/CLI change needed: add eu-west-3 & ap-south-1 region identifiers and new sku codes (h100-nvl,a10g-mid). I'll send a PR today. Mei (sales): Pricing? Customers expect delta vs A100. Sanaa (infra PM): Committed rates will be posted on the pricing page later today. Ballpark: H100 committed ~15% discount vs on-demand H100 list; A10G-mid positioned for cost-sensitive inference. Exact numbers in the email + pricing doc. Priya (support): Draft KB: https://docs.internal/dedicated/h100-mumbai-paris (WIP). I'll add FAQ about SLA, migration, and billing. Mei (sales): Any SLA changes for Dedicated with H100? Sanaa (infra PM): No SLA change — Dedicated SLA still applies. We'll publish an addendum note mentioning hardware types supported. Luca (sre): Action items for infra team: (1) update region registry (Javier), (2) prepare canary runbook and rollbacks (Luca), (3) update observability dashboards for new SKUs (ops), (4) post incident channel watch during rollout. Sanaa (infra PM): Action items (short): 1) SDK & console region/sku update (Javier), 2) Support KB & customer comms (Priya), 3) Sales one-pager + pricing FAQ (Mei), 4) Canary run + rollout ops (Luca/deploy-bot). Please ack here if you own an item. Luca (sre): ack Javier (capacity eng): ack — PR incoming. Priya (support): ack — KB coming in ~2h. Mei (sales): ack :party: Sanaa (infra PM): Thanks all — I'll post the perf matrix and pricing links into this thread when available. If you want early access for a customer pilot, DM me and we'll carve capacity out of the reserved pool.76:Tec5,Dedicated capacity recommendation & ops playbook for Sparkle Inc: H100/A100/L4 mix From: Ava Chen To: Emma Li Cc: Dev Patel , Lauren Bishop Date: Wed, 02 Dec 2026 09:12:00 -0800 Subject: Dedicated capacity recommendation & ops playbook for Sparkle Inc: H100/A100/L4 mix Hi Emma — great call this morning. As promised, I put together a recommendation and an ops playbook for Sparkle's Dedicated deployment that balances cost, latency SLOs, and reliability for the three primary workloads you called out: user-facing chat (short contexts, strict P95 latency), large-document summarization (long context, throughput-sensitive), and embeddings/batching (high throughput, lower latency requirements). High-level recommendation (summary): - Primary fleet: mix of L4 + H100 with a small A100 reserve. - L4 pool (40%): handle chat and embeddings at low cost and excellent token/sec for short contexts and embeddings throughput. - H100 pool (50%): serve long-context summarization and retrieval-augmented generation where memory and FP16 perf matter. - A100 reserve (10%): compatibility tier for heavy KV-cache / legacy quantized models and as an overflow for bursty long sequences. Why this mix? - Memory vs perf: H100 gives higher HBM capacity and matrix perf for long sequences and large batch KV cache. L4 is extremely cost-efficient for <8k contexts and embeddings. A100 sits in the middle and provides predictable FP16 throughput for some quantized variants you may want to pin. - Ops & reliability: we recommend pinning your mission-critical model to H100 with a fallback policy that routes to A100 or a smaller L4 ensemble when H100 capacity is saturated (graceful degradation on quality). Attached: a capacity model (sparkle_capacity_model_v2.xlsx) with projection by concurrency and expected qps, and a latency/throughput chart (sku_latency_tradeoffs.png). See the Drive link in-line if you want the live workbook. Next steps I suggest: 1) We spin up a small canary for your chat product on L4 (2 instances) and the summarization pipeline on H100 (1 instance) for 2 weeks. 2) Run the workload profiler (we'll share the script) across the canary to capture real seq-length distribution and prefix cache hit rates. 3) After profiling, we finalize exact node counts and reservation tiers (monthly commit) + define a burst pool for the A100 reserve. Operational notes: - Autoscaling: use CPU-meta autoscaling for the L4 pool to preserve latency; use queue-depth based scaling for H100 to maximize throughput. - KV cache: for long-context jobs pin to H100 to avoid cross-node KV cache thrash. - Cost control: set a per-route token cap and rate limits; Redwood Optimize can propose quantization for summarization that cuts cost up to ~30% with minimal quality drop (we can run evals). Emma — can you confirm the SLOs we should target (P95 latency in ms) for chat vs summarization and whether you'd like the A100 reserve to be billed as burst or reserved capacity? Once you confirm, we'll lock the canary start date. Thanks, Ava Chen Account Executive, Redwood Inference ava_chen@redwood.com +1 (415) 555-0199 -- Attachments: sparkle_capacity_model_v2.xlsx (spreadsheet), sku_latency_tradeoffs.png (image) From: Dev Patel To: Ava Chen Cc: Emma Li , Lauren Bishop Date: Wed, 02 Dec 2026 11:03:00 -0800 Subject: Re: Dedicated capacity recommendation & ops playbook for Sparkle Inc: H100/A100/L4 mix Thanks Ava — added a few technical clarifications after reviewing the profiler expectations and the projected concurrency in the spreadsheet. Quick notes for the team and for Emma: 77:Td76,eng-infra kevin: Quick check — we need to confirm GPU availability for the eu-central burst pool next week. Primary goal: H100s for latency-critical endpoints, fallback A100s. Vendor A says 12 H100s available next week, Vendor B can do 8 but with ~3w lead. Anyone seeing different quotas or constraints? marta: On our side we have 10 H100s already reserved in Dedicated for other pools. GCE console shows ~6 spare H100 quota in eu-central1 but that's not guaranteed. Pulling exact numbers now. ali: FYI procurement ticket #4521 has vendor-A proposal attached: https://procure.internal/4521 — if we want 20 H100 total we should start PO flow today. cc @marta procurement-bot: Ticket 4521 update: approval requested -> approved for 8 H100. PO created (PO-7789). ETA estimate: 7 business days. ria: Can we do a staggered plan? Use vendor-A H100 for the primary canary and then cloud preemptible A100s as overflow. Note: embedding endpoints require tight p95 latency — H100 preferred for critical slices. kevin: Our latency SLO is 50ms p95 for short sequences. A100 preemptible might increase p95. How many A100s could we add as fallback without violating SLO? marta: Rough calc: with small batching and KV-cache reuse, ~16 A100s could handle ~60% of peak, p95 would likely drift ~10-20ms. Suggest initial canary: 4 H100 + 6 A100, with strict routing and rollout checks. ali: I'll capture the canary plan in the ticket and call out finance deadline (vendor-B needs payment in 5 days). Canary snippet I'm adding here: ``` canary: region: eu-central1 primary: vendor-A H100 x4 fallback: cloud A100 x6 (preemptible) rollout: 25% -> 50% -> 100% over 4h monitor: p95_latency, error_rate, token_throughput rollback: immediate to previous model version + route to dedicated pool ``` kevin: @marta can you also confirm cloud quotas for A100 preemptible can hit 10 in the window 2026-03-25 10:00-14:00 CET? Need to schedule the canary slot. marta: Checking console + quota API now. GCE shows quota for a100-preemptible = 12 in eu-central1. Requesting increase would take ~48h, so current number should work for the canary-sized plan (6-10). Opening a support case to lock it down. procurement-bot: PO-7789 status: vendor confirmed invoice received. ETA revised to 5 business days. ria: Reminder: tag instances with routing label `latency-priority` and enable cache warming (pre-roll) 30m before rollout. Also ensure tracing headers are toggled on for canary traffic. ali: Added node labels in k8s plan and added a small rollout job: `kubectl label nodes --selector=pool=eu-burst latency-priority=true` and prewarm job: `kubectl run prewarm --image=redwood/prewarm:stable --replicas=4 -- /warm --regions=eu-central1` kevin: Great — I'll schedule the canary for 2026-03-26 09:00 CET (canary window 09:00–13:00). If quotas or vendor ETA change, ping this thread. :rocket: marta: Also pinged cloud-sales and opened support case CS-9012 for quota confirmation. They asked for the support case — response ETA 24–48h. cloud-sales: Thanks for the ping, @marta — we've flagged CS-9012 for expedited review. If you share the exact project ID and expected usage spike we can try to lock capacity for the window. kevin: I'll drop project ID and expected usage in CS-9012. TL;DR: canary plan ready, PO in flight, scheduling for 2026-03-26 09:00 CET. Will update if anything breaks. marta: ack. ali: ack. ria: ack :thumbsup:78:T9c3,eng-infra morgan: Quick ask — can we secure a small H100 slot in apac (asia-southeast1) next Wed 09:00-12:00 UTC for a customer canary? Need 1 node (80GB) for ~3 hours. cc @gleb-vendor. Also will need a brief terraform PR + smoke window. What is lead time and ballpark cost? gleb-vendor: We do have very limited H100 capacity in asia-southeast1 that week. I can put 1 node in standby/reserve for you but need billing sign-off. Ballpark: ~ $180/hr for a dedicated node (before discounts). If you can provide project id + billing owner I can start the reserve. sam: To add — if you want a dedicated standby we'll need a quota bump on the project: redwood-canary-apac. @jess-finance will need to approve cost and we need a terraform change to create the pool allocation (small snippet below). ```hcl resource "redwood_gpu_pool" "canary_apac_h100" { region = "asia-southeast1" sku = "h100-80gb" min_nodes = 1 max_nodes = 1 purpose = "customer-canary-2026-03" } ``` anya: I can open the infra PR today with that config and a small job to run the canary workload. Will tag riley-oncall for the smoke. PR draft: https://github.com/redwood/infrastructure/pull/1234 (will switch to real when ready). riley-oncall: I can own the oncall window. Suggest we do a quick staging smoke at 03:00 UTC the same day to validate networking and KV cache warming before the customer run at 09:00. Also let's enable verbose logging for the first 10 minutes. gleb-vendor: Sounds good — I'll hold the slot provisionally for 24h pending billing. Ping me when finance confirms and when the PR is merged so we can attach the tenant to the allocation. jess-finance: Provisional OK up to $600 for the run. Need exact estimate and project code to finalize chargeback. Please include expected token throughput if you want a more precise cost projection. morgan: Thanks — will send exact estimate and token throughput by EOD today. Project: redwood-canary-apac. I'll merge the PR after anya's finalizes and will paste the PR link here. :thumbsup: deploy-bot: terraform plan for redwood-canary-apac succeeded in CI (job 9876). Plan shows +1 gpu pool resource pending apply. anya: PR updated, added smoke job and logging flags. Ready for review — tagging @sam @riley-oncall. sam: LGTM, please include a short rollback playbook in the PR description (how to drain and fall back to hosted model variant). morgan: Will add rollback steps and then merge once finance confirms final number. Thanks everyone!79:Tf9f,Regional capacity forecast & rollout timeline Q4-2025 to Q2-2026 Purpose: Short regional capacity forecast across NA / EU / APAC to inform Dedicated and Hosted capacity commitments and Private deployment timelines. Focus window: 2025-Q4 through 2026-Q2. This is a working sheet: numbers are estimates, callouts list assumptions and open questions. Executive summary: - Net add of capacity across regions driven by hosted demand and two enterprise Reserved Dedicated deals forecast in Q1 2026. - Prioritize NA (us-east/us-west) for low-latency hosting and Dedicated pools; EU follows for compliance and capacity isolation; APAC ramp will be staged to avoid overspend and to validate H100 performance in region. - Target steady-state utilization: 62% (hosted baseline) with 20% headroom buffer for burst and canary rollouts. Assumptions (high level): - Typical model mix: 65% medium LLMs (8-20B), 30% large LLMs (70B-180B), 5% embeddings/other. - Instance types planned: A100 80GB and H100 80GB (H100 trial in EU and APAC before full rollout). - Cost assumptions (cloud list price estimate): A100-hour = $3.20, H100-hour = $7.20 (spot savings not guaranteed for Dedicated). - Average tokens/sec per GPU estimated from current benchmarks: medium models ~220 tok/s, large models ~70 tok/s. - We maintain 20% spare capacity for regional failover and autoscaling headroom. Capacity table (planned active GPUs by month, rolling forecast) Format: Month | NA (A100/H100) | EU (A100/H100) | APAC (A100/H100) | Notes 2025-10 | 240/0 | 120/0 | 40/0 | baseline (actual) 2025-11 | 260/8 | 130/0 | 45/0 | H100 pilot hardware procurement begins (NA) 2025-12 | 300/16 | 140/4 | 50/0 | holiday season headroom, small EU H100 pilot 2026-01 | 360/32 | 160/8 | 60/4 | Dedicated commit1 expected (NA) -> +40 GPUs net (reserved A100-equivalent) 2026-02 | 420/48 | 180/16 | 75/8 | scale for growth + enterprise onboarding; APAC H100 trial starts in region 2026-03 | 480/64 | 220/24 | 100/16 | Pre-rollout for Dedicated commit2 (EU Dedicated lead) 2026-04 | 540/80 | 280/32 | 140/24 | Post-Dedicated2 full provisioning begins (mix A100/H100) 2026-05 | 600/96 | 340/40 | 180/32 | Steady ramp; autoscaling tuning in NA and EU 2026-06 | 660/112 | 400/48 | 220/40 | Target capacity for Q2 closeout; review utilization targets Target utilization and economics (monthly averages) - NA target utilization: 65% peak, 55% off-peak. Effective utilization after batching/caching: +8–12% token throughput improvement. - EU target utilization: 62% peak (allowing for compliance routing), 50% off-peak. - APAC target utilization: 48–55% initially during H100 validation, target 60% after Q2. - Expected blended cost per effective GPU-hour (after batching gains and spot where possible): NA $2.45, EU $2.90, APAC $3.10. Rollout timeline & milestones (by region) - NA (priority): - Nov 2025: H100 pilot on one cluster (internal performance validation). - Jan 2026: Provision Dedicated commit1 (NA reserved pool); finalize autoscaler policies for steady throughput. - Mar 2026: Canary live for new latency-sensitive routes; monitor tail-latency and KV cache hit rates. - EU: - Dec 2025: Small H100 pilot (privacy compliance testing). - Mar 2026: Dedicated offer discussions with EU enterprise; readiness planning for Dedicated cluster provisioning. - Apr 2026: Start dedicated provisioning for commit2; route EU traffic to new pools with canary policies. - APAC: - Feb 2026: H100 validation in APAC (short trial window). - May 2026: Gradual scale-up if H100 perf + cost acceptable; otherwise fall back to A100-heavy mix. Operational notes and dependencies: - Rack/cabinet provisioning not applicable for cloud Dedicated but inventory lead times for bare-metal partners estimated at 6–10 weeks. - Network: ensure cross-region egress cost accounting and inter-region fallback policy to avoid surprises. - Observability: need RL metrics for per-GPU token throughput, KV cache hit rates, and per-route latency SLOs. 7a:Tf8b,2) Ben: pricing/commitment tiers and the guardrails on burst > 2x - are standbys automatically billed, or do we require pre-approval? Need the commercial language we can send to procurement. 3) Grace: confirm operational lead time to add 10 H100-equivalent slots in a dedicated pool and the steps for hosted fallback routing (networking, latency delta). I want to keep the external reply concise — just numbers and a phased proposal. Hit reply-all with your inputs by EOD so I can consolidate. -Kim From: Marcus Lin To: Kimberly Park , Ben Carter , Grace O'Connor Date: 2026-11-22T14:40:00-08:00 Subject: Re: FW: Dedicated capacity questionnaire (internal: OpticAI RFI) Kim, See numbers below (rounded and conservative): - H100-equivalent (80GB): medium chat (seq=128): ~1,200 tokens/sec; long-gen (seq=2048): ~160 tokens/sec. - A100-equivalent (80GB): medium chat: ~850 tokens/sec; long-gen: ~110 tokens/sec. - L4 class: medium chat: ~2200 tokens/sec (small batch, short seq); long-gen: ~N/A (not recommended for >1024). Recommended pool for 2M tokens/month (99th pct tail target 600ms): - Pilot (30 days): 2x H100 + 4x A100 + L4 cache tier (2 nodes). - Phase 2 (50% commit): 4x H100 + 6x A100 + L4 cache tier (4 nodes). - Full commit (100%): 6x H100 + 10x A100 with autoscale headroom and reserved L4 cache. Burst/fallback: we can promise 2.5x burst for up to 72 hours with pre-negotiated credit allowances. After that, fallback to hosted pool with cross-region routing (expected latency increase 25-40ms RTT). I uploaded the spreadsheet draft as gpu_mix_recommendation.xlsx (link in internal drive). If OpticAI can confirm pct of short chats vs long-gen and concurrency profiles we can tighten numbers. -Marcus From: Kimberly Park To: Samira Qureshi , Ethan Lo Cc: Ben Carter , Marcus Lin Date: 2026-11-23T09:22:00-08:00 Subject: Re: Dedicated capacity questionnaire (OpticAI RFI) Hi Samira / Ethan — Thanks for the RFI. Please find a consolidated reply below and attached short workbook for your procurement team. Key points up front: - Supported SKUs and recommended mix: H100-equivalent (primary for large seq), A100-equivalent (balanced), L4 cache tier (short, high-req throughput). See attached table for tokens/sec per SKU. - Throughput estimates (examples): H100: 1,200 t/s (seq=128), 160 t/s (seq=2048). A100: 850 t/s / 110 t/s. L4: 2,200 t/s (short seq). These are conservative per-slot expectations to give headroom. - Phased reservation plan (mutual action plan): * Pilot (30 days): 2x H100 + 4x A100 + L4 cache (2) — validate tail latency and cache hit rates. * Phase 2 (50% commit): 4x H100 + 6x A100 + L4 (4) — run steady-state traffic for 60 days. * Full (100% commit): 6x H100 + 10x A100 with automatic autoscale headroom and agreed burst credits. - Burst and fallback: Up to 2.5x burst for short windows (72 hours) via reserved credit pool. If sustained beyond agreed burst, requests route to Redwood hosted pool (transparent fallback) with an expected median latency increase of 25–40ms; we provide monitoring and automatic retries. You can opt into a High-Priority lane (paid) to avoid fallback. - SLA language highlights: Dedicated reservation covers committed throughput and median latency. 99th pct tail latency is included in SLO monitoring; we offer an addendum with deeper guarantees for an additional commitment. Eviction: preemption only under extreme platform-wide events; otherwise slice isolation is enforced. Next steps: - Can you share a representative token-length histogram and concurrency profile (simple CSV) so we can run the workbook and produce a final price/slot recommendation? - Proposed meeting to review the workbook: Wed 11/24 10:00am PT or Thu 11/25 2:00pm PT — which works? 7b:Tf66,- Burst to 16 GPUs: what is the cold-start time to spin from baseline to burst capacity for H100? We need to understand pre-warm vs spin-up latency — for live triage flow we can't have >1s tail growth. - Multi-region: do you recommend active-active or active-passive for our use case? We expect East coast traffic to be ~65% at launch. - Any limits on model context windows that meaningfully change the GPU counts (we plan to run 32k-context retrieval-augmented generation models). Also, procurement will want the exact SOW line items for "warm pool" and the PL/PO language for credits if provisioning delays occur (we have an internal date we can't miss). If you can answer these, we can loop procurement and greenlight Pilot or Growth. —Rajiv CTO, Saturn Health rajiv.mehta@saturnhealth.com On Tue, Oct 20, 2026 at 09:14 AM Karthik Iyer wrote: > [original message trimmed for thread] From: Marissa Cole To: Rajiv Mehta , Anna Park Cc: Karthik Iyer , diego_fernandez@redwood.ai Date: Wed, 21 Oct 2026 10:40:00 -07:00 Subject: Re: Cluster sizing & scaling path for Saturn Health dedicated capacity Rajiv / Anna — looping in a couple of technical specifics to address the questions below and make the SOW lines explicit. 1) Billing / min-guaranteed vs reserved baseline - When you sign a reserved baseline (e.g., 8 GPUs), billing is at the reserved baseline rate. "Min-guaranteed" reflects the SLAs on how many hosts stay online during rolling maintenance or host failures; it is not a separate billable floor. If you reserve 8 but we guarantee 6 during maintenance windows, you still pay the reserved 8 monthly commitment. We can add a clause for credits if we miss agreed availability targets. 2) Burst/scale-up timing and warm pool - Cold spin-up for physical GPUs (if provisioning new hosts) can be 3–5 minutes; our autoscaler uses a warm pool to reduce that to <10s for most workloads. Warm pool is effectively idle pre-warmed instances (partial KV cache warm) — recommended size: ~10% of baseline (we included line item). - For H100 with our KV cache and prefix caching, tail latency growth is typically under 200–300ms when a warm pool is sized correctly. If you need <100ms tails we recommend >15% warm pool and more aggressive KV pre-population. 3) Multi-region recommendation - For your profile (65% East), we recommend active-primary (US-East) + standby (US-West) active-passive with cross-region failover for cost efficiency. Active-active makes sense if you plan high geo-traffic or strict locality; it increases baseline GPUs because you must reserve capacity in both. 4) Context windows and GPU math - 32k-context models increase KV cache and memory pressure. For H100-80GB you can comfortably run a moderate throughput serving node with 2–3 concurrent 32k calls depending on model size. We assumed 2 concurrent 32k calls per GPU for the capacity workbook. If you expect higher concurrency we should move the per-GPU concurrency assumption down (e.g., 1.2–1.5) which will raise GPU counts. Attached: capacity-plan.xlsx (breaks down per-model memory, assumed concurrency, and warm-pool sizing). I've also updated the workbook with explicit SOW language for warm pool, crediting for provisioning delays, and burst billing terms. If this looks good I can draft the SOW language and hand it to Rishi for pricing & credit terms. —Marissa Principal Infra Engineer, Redwood Inference marissa_cole@redwood.ai ----- Forwarded message ----- From: Karthik Iyer Date: Tue, 20 Oct 2026 09:14:00 -07:00 Subject: Cluster sizing & scaling path for Saturn Health dedicated capacity To: Rajiv Mehta , Anna Park [original attachments omitted] From: Rishi Malhotra 7c:T681,- We recommend sizing pools to expected p50 peak + 1.5x burst cushion; we’ll walk through sample math per workload profile. - Autoscale lag: warmup of weighty models can be 20–45s depending on quantization + kernel warm paths; discuss warm pool sizing to hit P99 latency SLOs. - Isolation: dedicated pools per tenant for high isolation vs. soft multi-tenant pools with cgroup/namespace quotas for cost efficiency. (attachment: proposed-canary-plan.xlsx) —Marcus Marcus Lin Principal Engineer, Serving Runtime marcus_lin@redwood.com From: laura_bennett To: Amit Desai , Marcus Lin , Evelyn Hart Date: 2025-11-26T14:02:00-08:00 Subject: Confirmed: Wed 11/26 10:00–11:00 PT (GPU capacity architecture review) — calendar updated All — Confirmed for Wed 2025-11-26 10:00–11:00 PT / 13:00–14:00 ET. I updated the invite and included a video link and read-only copy of the brief in the calendar. Please find pre-reads attached in the event and on Drive: https://drive.google.com/open?id=1A2b3C4gpu-arch. Quick housekeeping: - Please add any specific workload traces you want us to look at to the Drive folder before EOD Tuesday so Marcus can bring numbers. - Marcus, Evelyn — can you prepare a 5–10 minute runbook sketch for canaries/fallbacks? - Amit/Priya — let us know if there are any compliance constraints (network isolation/KMS) you want covered. Looking forward to it. Laura Bennett Enterprise AE, Redwood Inference laura_bennett@redwood.com (Updated calendar event + attachments: redwood-gpu-arch-brief-v1.pdf, proposed-canary-plan.xlsx) 0:{"P":null,"c":["",""],"q":"","i":false,"f":[[["",{"children":["__PAGE__",{},"$undefined","$undefined",4608]},"$undefined","$undefined",4624],[["$","$1","c",{"children":[[["$","link","0",{"rel":"stylesheet","href":"/_next/static/immutable/chunks/2wtply62jadr_.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}],["$","script","script-0",{"src":"/_next/static/immutable/chunks/2corl6wyj4669.js","async":true,"nonce":"$undefined"}]],["$","html",null,{"lang":"en","className":"lato_aa86c0ae-module__7bD5Rq__variable geist_mono_3342dced-module__ksAy2q__variable h-full","data-theme":"light","suppressHydrationWarning":true,"children":["$","body",null,{"className":"min-h-full","children":["$","$L2",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L3",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":[[["$","title",null,{"children":"404: This page could not be found."}],["$","div",null,{"style":{"fontFamily":"system-ui,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif,\"Apple Color Emoji\",\"Segoe UI Emoji\"","height":"100vh","textAlign":"center","display":"flex","flexDirection":"column","alignItems":"center","justifyContent":"center"},"children":["$","div",null,{"children":[["$","style",null,{"dangerouslySetInnerHTML":{"__html":"body{color:#000;background:#fff;margin:0}.next-error-h1{border-right:1px solid rgba(0,0,0,.3)}@media (prefers-color-scheme:dark){body{color:#fff;background:#000}.next-error-h1{border-right:1px solid rgba(255,255,255,.3)}}"}}],["$","h1",null,{"className":"next-error-h1","style":{"display":"inline-block","margin":"0 20px 0 0","padding":"0 23px 0 0","fontSize":24,"fontWeight":500,"verticalAlign":"top","lineHeight":"49px"},"children":404}],["$","div",null,{"style":{"display":"inline-block"},"children":["$","h2",null,{"style":{"fontSize":14,"fontWeight":400,"lineHeight":"49px","margin":0},"children":"This page could not be found."}]}]]}]}]],[]],"forbidden":"$undefined","unauthorized":"$undefined"}]}]}]]}],{"children":[["$","$1","c",{"children":[["$","$L4",null,{"prepared":{"proxima-429":{"answer":"$5","sources":[{"key":"SRC:0002","kind":"document","fileName":"Proxima Bank unexpected spike in 429s after enterprise priority routing rollout (chat embeddings).txt","score":1,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Jira","page":1,"fileId":"01a09931-0ed6-7741-8867-c72c68a870d7","text":"$6","textOnly":true},{"key":"SRC:0004","kind":"document","fileName":"Hosted API SLOs Enterprise Route Tiers (Hot-route capacity protection).txt","score":1,"isCited":true,"n":2,"localUrl":null,"liveUrl":null,"system":"Wiki","page":0,"fileId":"01a09931-3830-75a6-8a53-e406eaa58a62","text":"$7","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Hosted API SLOs Enterprise Route Tiers (Hot-route capacity protection).txt","score":1,"isCited":true,"n":3,"localUrl":null,"liveUrl":null,"system":"Wiki","page":1,"fileId":"01a09931-3830-75a6-8a53-e406eaa58a62","text":"$8","textOnly":true},{"key":"SRC:0003","kind":"document","fileName":"Proxima Bank unexpected spike in 429s after enterprise priority routing rollout (chat embeddings).txt","score":1,"isCited":true,"n":4,"localUrl":null,"liveUrl":null,"system":"Jira","page":2,"fileId":"01a09931-0ed6-7741-8867-c72c68a870d7","text":"2026-01-23 09:05 PT - Rafael Mendes (SRE): Post-change review: no negative impact on enterprise SLO dashboards; Proxima-specific 429s reduced; reason mix shifted mostly to quota_exceeded at end-of-day usage peaks (expected). Marking ticket resolved with note: longer-term solution may be Dedicated capacity or revisiting contracted burst terms.\nResolved with (1) customer education on 429 reason codes and correct retry/backoff behavior, and (2) a time-boxed, region-scoped burst exception for PROXIMA-ENT-014 on chat+embeddings (14-day expiry) with automated rollback guardrails. Customer to deploy improved client-side exponential backoff + jitter and concurrency caps. Separate guidance provided for quota_exceeded events (smooth usage / avoid end-of-window spikes).","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Proxima Bank unexpected spike in 429s after enterprise priority routing rollout (chat embeddings).txt","score":1,"isCited":false,"n":5,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c72c68a870d7","text":"$9","textOnly":true}],"secs":36.5,"effort":"ultra","score":991},"dek-conflict":{"answer":"$a","sources":[{"key":"SRC:0017","kind":"document","fileName":"Key Management Standard KMS Integrations.txt","score":0.7903234727372634,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Wiki","page":2,"fileId":"01a09931-3830-75a6-8a53-e41bf225fa9c","text":"$b","textOnly":true},{"key":"SRC:0019","kind":"image","fileName":"Key rotation - dual-read dual-decrypt window (ADR-008).jpg","score":0.9748264698305706,"isCited":true,"n":1,"localUrl":"/media/diagrams/Key rotation - dual-read dual-decrypt window (ADR-008).png","liveUrl":"https://storage.googleapis.com/graphon-private-upload-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0993f-ebed-774e-a017-b97cdd91802f/Key%20rotation%20-%20dual-read%20dual-decrypt%20window%20%28ADR-008%29.jpg?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T120905Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Diagrams"},{"key":"SRC:0012","kind":"document","fileName":"ADR-007 Standard KMS Client Wrapper and Failure Policy.txt","score":1.0122038125991821,"isCited":true,"n":2,"localUrl":null,"liveUrl":null,"system":"Wiki","page":2,"fileId":"01a09931-3830-75a6-8a53-e3bfb0523536","text":"$c","textOnly":true},{"key":"SRC:0016","kind":"document","fileName":"Key Management Standard KMS Integrations.txt","score":0.7615595770120318,"isCited":true,"n":3,"localUrl":null,"liveUrl":null,"system":"Wiki","page":1,"fileId":"01a09931-3830-75a6-8a53-e41bf225fa9c","text":"$d","textOnly":true},{"key":"SRC:0015","kind":"document","fileName":"Key Management Standard KMS Integrations.txt","score":0.8417096362012099,"isCited":true,"n":4,"localUrl":null,"liveUrl":null,"system":"Wiki","page":0,"fileId":"01a09931-3830-75a6-8a53-e41bf225fa9c","text":"$e","textOnly":true},{"key":"SRC:0018","kind":"document","fileName":"Key Management Standard KMS Integrations.txt","score":0.7756505642428975,"isCited":true,"n":5,"localUrl":null,"liveUrl":null,"system":"Wiki","page":3,"fileId":"01a09931-3830-75a6-8a53-e41bf225fa9c","text":"$f","textOnly":true},{"key":"SRC:0003","kind":"document","fileName":"Acme Bank security review KMS rotation materials outage behavior (customer-facing).txt","score":1.0718221089786926,"isCited":true,"n":6,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-02ab6eb68123","text":"$10","textOnly":true},{"key":"SRC:0004","kind":"document","fileName":"Acme Bank security review KMS rotation materials outage behavior (customer-facing).txt","score":1.022973661351615,"isCited":true,"n":7,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed4-7228-bdd7-02ab6eb68123","text":"$11","textOnly":true},{"key":"SRC:0014","kind":"document","fileName":"ADR-008 Key rotation strategy (key versioning, dual-read dual-decrypt window, cutover rollback).txt","score":0.7935685060643333,"isCited":true,"n":8,"localUrl":null,"liveUrl":null,"system":"Wiki","page":0,"fileId":"01a09931-3830-75a6-8a53-e3c0a7684974","text":"$12","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Acme Bank security review KMS rotation materials outage behavior (customer-facing).txt","score":0.9158980435437686,"isCited":true,"n":9,"localUrl":null,"liveUrl":null,"system":"Email","page":2,"fileId":"01a09931-0ed4-7228-bdd7-02ab6eb68123","text":"$13","textOnly":true},{"key":"SRC:0011","kind":"document","fileName":"ADR-007 Standard KMS Client Wrapper and Failure Policy.txt","score":0.8430092286574595,"isCited":true,"n":10,"localUrl":null,"liveUrl":null,"system":"Wiki","page":1,"fileId":"01a09931-3830-75a6-8a53-e3bfb0523536","text":"$14","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Crypto hierarchy, rotation, and SIEM export runbook.txt","score":0.7680313142174993,"isCited":false,"n":11,"localUrl":null,"liveUrl":null,"system":"Drive","page":0,"fileId":"01a09931-0ed3-730a-8f12-e2e070a67f83","text":"$15","textOnly":true},{"key":"SRC:0002","kind":"document","fileName":"SOC2 ISO control translation 1 1 dry-run scratchpad.txt","score":0.7649367377247691,"isCited":false,"n":12,"localUrl":null,"liveUrl":null,"system":"Drive","page":1,"fileId":"01a09931-0ed3-730a-8f12-e360b377085e","text":"$16","textOnly":true},{"key":"SRC:0006","kind":"document","fileName":"Re HealthBridge KMS BYOK transient custody, emergency rotation & TLS profile.txt","score":0.7982338142813182,"isCited":false,"n":13,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed5-7746-ba56-647b653c5a14","text":"$17","textOnly":true},{"key":"SRC:0007","kind":"document","fileName":"ENG 11097 priv cross tenant audit suppression and kms rotation fix.txt","score":0.777542065375675,"isCited":false,"n":14,"localUrl":null,"liveUrl":null,"system":"Linear","page":1,"fileId":"01a09931-0ed6-7741-8867-c7ccb5e1067d","text":"$18","textOnly":true},{"key":"SRC:0008","kind":"document","fileName":"Vendor risk questionnaire prep - Lumos Health.txt","score":0.7792346731519134,"isCited":false,"n":15,"localUrl":null,"liveUrl":null,"system":"Meetings","page":1,"fileId":"01a09931-236b-7265-918a-68974808a9a8","text":"$19","textOnly":true},{"key":"SRC:0009","kind":"document","fileName":"#eng-security - 1799123456 auditor kms hsm access followup.txt","score":0.778011345658898,"isCited":false,"n":16,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-236e-7731-8353-1cabfb6621ab","text":"$1a","textOnly":true},{"key":"SRC:0010","kind":"document","fileName":"ADR-007 Standard KMS Client Wrapper and Failure Policy.txt","score":0.8229975606785018,"isCited":false,"n":17,"localUrl":null,"liveUrl":null,"system":"Wiki","page":0,"fileId":"01a09931-3830-75a6-8a53-e3bfb0523536","text":"$1b","textOnly":true},{"key":"SRC:0013","kind":"document","fileName":"ADR-007 Standard KMS Client Wrapper and Failure Policy.txt","score":0.7982942651410019,"isCited":false,"n":18,"localUrl":null,"liveUrl":null,"system":"Wiki","page":3,"fileId":"01a09931-3830-75a6-8a53-e3bfb0523536","text":"$1c","textOnly":true},{"key":"SRC:0020","kind":"image","fileName":"Key rotation - dual-read dual-decrypt window (ADR-008).jpg","score":0.9860102246068482,"isCited":false,"n":2,"localUrl":"/media/diagrams/Key rotation - dual-read dual-decrypt window (ADR-008).png","liveUrl":"https://storage.googleapis.com/graphon-private-upload-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-601e2d7ab1a2/Key%20rotation%20-%20dual-read%20dual-decrypt%20window%20%28ADR-008%29.jpg?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T120905Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Diagrams"}],"secs":14.8,"effort":"standard","score":1640},"token-discrepancy":{"answer":"The provided sources do not give a channel-by-channel count, so they do not establish which intake channel has the most reports. They identify the tracked origins as Jira, Slack, and Gmail, and specify that weekly reporting must count each case by its original channel—not by later-created support-ticket artifacts. [[SRC:0020]]","sources":[{"key":"SRC:0020","kind":"document","fileName":"Support Ops Reconciling token accounting discrepancy reports (intake triage).txt","score":1.2006022989929415,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Wiki","page":0,"fileId":"01a09931-3830-75a6-8a53-e4972da5824c","text":"$1d","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Usage meter audit Feb 2026 (working notes).txt","score":0.6682305517332083,"isCited":false,"n":2,"localUrl":null,"liveUrl":null,"system":"Drive","page":1,"fileId":"01a09931-0ed3-730a-8f12-e37c3d45a47f","text":"$1e","textOnly":true},{"key":"SRC:0002","kind":"document","fileName":"Counting glitch impacting April invoices next steps & timeline.txt","score":0.6719892044671137,"isCited":false,"n":3,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-032210b3b977","text":"$1f","textOnly":true},{"key":"SRC:0003","kind":"document","fileName":"Fwd Question about token accounting discrepancy (Feb invoice vs usage).txt","score":0.8505249278034068,"isCited":false,"n":4,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-0392c970b567","text":"$20","textOnly":true},{"key":"SRC:0004","kind":"document","fileName":"Phantom token counts May 21-22 trace, compensation, and evidence request.txt","score":0.8889881751680189,"isCited":false,"n":5,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed5-7746-ba56-640598d5af11","text":"$21","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Phantom token counts May 21-22 trace, compensation, and evidence request.txt","score":0.7098719487220245,"isCited":false,"n":6,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed5-7746-ba56-640598d5af11","text":"$22","textOnly":true},{"key":"SRC:0006","kind":"document","fileName":"Fix double-counting on stream retry fallback for Hosted API metering.txt","score":0.6899456900142553,"isCited":false,"n":7,"localUrl":null,"liveUrl":null,"system":"GitHub","page":0,"fileId":"01a09931-0ed5-7746-ba56-655ec71b12f8","text":"$23","textOnly":true},{"key":"SRC:0007","kind":"document","fileName":"Acme AI reports being double-charged tokens on retry fallback path (streaming) in Hosted API.txt","score":0.7414264816432592,"isCited":false,"n":8,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c6db16c2cb28","text":"$24","textOnly":true},{"key":"SRC:0008","kind":"document","fileName":"Acme AI reports being double-charged tokens on retry fallback path (streaming) in Hosted API.txt","score":0.7173684427360564,"isCited":false,"n":9,"localUrl":null,"liveUrl":null,"system":"Jira","page":1,"fileId":"01a09931-0ed6-7741-8867-c6db16c2cb28","text":"$25","textOnly":true},{"key":"SRC:0009","kind":"document","fileName":"SUP 18510 streaming throughput metric mismatch.txt","score":0.7117507363435768,"isCited":false,"n":10,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c73f596b2692","text":"$26","textOnly":true},{"key":"SRC:0010","kind":"document","fileName":"SUP 18510 streaming throughput metric mismatch.txt","score":0.7096752830421162,"isCited":false,"n":11,"localUrl":null,"liveUrl":null,"system":"Jira","page":1,"fileId":"01a09931-0ed6-7741-8867-c73f596b2692","text":"$27","textOnly":true},{"key":"SRC:0011","kind":"document","fileName":"SUP 28451 invoice tokens do not match usage api.txt","score":0.7950184662632471,"isCited":false,"n":12,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c74379fa526b","text":"$28","textOnly":true},{"key":"SRC:0012","kind":"document","fileName":"SUP 28451 invoice tokens do not match usage api.txt","score":0.7704582764313557,"isCited":false,"n":13,"localUrl":null,"liveUrl":null,"system":"Jira","page":1,"fileId":"01a09931-0ed6-7741-8867-c74379fa526b","text":"$29","textOnly":true},{"key":"SRC:0013","kind":"document","fileName":"token accounting discrepancy between streaming usage and invoice export.txt","score":0.863560714114243,"isCited":false,"n":14,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c7706c2721b4","text":"$2a","textOnly":true},{"key":"SRC:0014","kind":"document","fileName":"#eng-platform - 2109876543 rate limiter toolcall metering clarify.txt","score":0.7016100827366218,"isCited":false,"n":15,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-236d-7700-af57-aae0da63515f","text":"$2b","textOnly":true},{"key":"SRC:0015","kind":"document","fileName":"#finance - 1750100000 tokenizer migration recount ask.txt","score":0.8602222199962756,"isCited":false,"n":16,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-236e-7731-8353-1cf264bb33e9","text":"$2c","textOnly":true},{"key":"SRC:0016","kind":"document","fileName":"#support - 1723849205.txt","score":0.983609676361084,"isCited":false,"n":17,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-3830-75a6-8a53-e33cf3f51549","text":"$2d","textOnly":true},{"key":"SRC:0017","kind":"document","fileName":"#support - 2156789012 late ack handoff token surge.txt","score":0.7274554255609657,"isCited":false,"n":18,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-3830-75a6-8a53-e37c5624b803","text":"$2e","textOnly":true},{"key":"SRC:0018","kind":"document","fileName":"Billing Meter Spec Hosted API Tokens & Requests.txt","score":0.6761504178899584,"isCited":false,"n":19,"localUrl":null,"liveUrl":null,"system":"Wiki","page":2,"fileId":"01a09931-3830-75a6-8a53-e3d2df416137","text":"$2f","textOnly":true},{"key":"SRC:0019","kind":"document","fileName":"P1 Incident Postmortem Token accounting mismatch in usage exports.txt","score":0.7305652302042063,"isCited":false,"n":20,"localUrl":null,"liveUrl":null,"system":"Wiki","page":0,"fileId":"01a09931-3830-75a6-8a53-e436ae6c4afc","text":"$30","textOnly":true}],"secs":7.3,"effort":"standard","score":558},"savings-mismatch":{"answer":"$31","sources":[{"key":"SRC:0004","kind":"document","fileName":"Optimization Suggestions estimated savings does not match billing usage totals (Arkadia Finance).txt","score":0.9523511520027145,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Jira","page":1,"fileId":"01a09931-0ed6-7741-8867-c7206f95a1b5","text":"$32","textOnly":true},{"key":"SRC:0003","kind":"document","fileName":"Optimization Suggestions estimated savings does not match billing usage totals (Arkadia Finance).txt","score":1.1540195896023682,"isCited":true,"n":2,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c7206f95a1b5","text":"$33","textOnly":true},{"key":"SRC:0007","kind":"document","fileName":"#support - 1771741044 beta customer feedback estimates seem off.txt","score":0.928699149710101,"isCited":true,"n":3,"localUrl":null,"liveUrl":null,"system":"Slack","page":1,"fileId":"01a09931-3830-75a6-8a53-e3556a546a4c","text":"$34","textOnly":true},{"key":"SRC:0002","kind":"document","fileName":"Savings estimator v0 for batching caching suggestions (cost latency deltas).txt","score":0.895068713343528,"isCited":true,"n":4,"localUrl":null,"liveUrl":null,"system":"GitHub","page":0,"fileId":"01a09931-0ed6-7741-8867-c6a4fe3cbff1","text":"$35","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Post-demo Prefetch prefix caching assumptions target 1M tokens.txt","score":0.7904617440988672,"isCited":false,"n":5,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed5-7746-ba56-6429f734f6f3","text":"$36","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Optimization Suggestions estimated savings does not match billing usage totals (Arkadia Finance).txt","score":0.8162966731561428,"isCited":false,"n":6,"localUrl":null,"liveUrl":null,"system":"Jira","page":2,"fileId":"01a09931-0ed6-7741-8867-c7206f95a1b5","text":"Provided reconciliation guidance and validated that observed savings aligns with estimate after accounting for (1) compute-cost vs net billing totals, (2) partial rollout percentage, (3) cache warm-up and baseline window alignment. No platform defect identified; follow-up request logged to improve UI copy around assumptions and to clarify savings are compute-cost estimates.","textOnly":true},{"key":"SRC:0006","kind":"document","fileName":"#support - 1771741044 beta customer feedback estimates seem off.txt","score":1.1040939688682634,"isCited":false,"n":7,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-3830-75a6-8a53-e3556a546a4c","text":"$37","textOnly":true},{"key":"SRC:0008","kind":"video","fileName":"Arm platform review - AmpereOne and Altra Max motherboards.mp4","score":0.8530676468245165,"isCited":false,"n":1,"localUrl":"/media/recordings/Arm platform review - AmpereOne and Altra Max motherboards.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-601fc4c98005/segments/segment_chunk_000000_000006.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121009Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":30,"end":35,"transcript":"$38"}],"secs":17.8,"effort":"standard","score":1397},"fallback-routing":{"answer":"$39","sources":[{"key":"SRC:0003","kind":"document","fileName":"Smart Routing fallback model region when overload breaker is open (bounded loop-safe).txt","score":1.1117523577287136,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"GitHub","page":0,"fileId":"01a09931-0ed6-7741-8867-c6a557db8e63","text":"$3a","textOnly":true},{"key":"SRC:0004","kind":"document","fileName":"Smart Routing fallback model region when overload breaker is open (bounded loop-safe).txt","score":0.9289430846136439,"isCited":true,"n":2,"localUrl":null,"liveUrl":null,"system":"GitHub","page":1,"fileId":"01a09931-0ed6-7741-8867-c6a557db8e63","text":"$3b","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Fallback model not triggering during overload (customer receives 503 OVERLOADED instead of routed fallback).txt","score":1.097523954584485,"isCited":true,"n":3,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c6fa90702991","text":"$3c","textOnly":true},{"key":"SRC:0006","kind":"document","fileName":"Fallback model not triggering during overload (customer receives 503 OVERLOADED instead of routed fallback).txt","score":0.9331830562342289,"isCited":true,"n":4,"localUrl":null,"liveUrl":null,"system":"Jira","page":1,"fileId":"01a09931-0ed6-7741-8867-c6fa90702991","text":"$3d","textOnly":true},{"key":"SRC:0007","kind":"document","fileName":"Fallback model not triggering during overload (customer receives 503 OVERLOADED instead of routed fallback).txt","score":0.8530317208236586,"isCited":true,"n":5,"localUrl":null,"liveUrl":null,"system":"Jira","page":2,"fileId":"01a09931-0ed6-7741-8867-c6fa90702991","text":"2025-12-13 08:40 PT — Peter Holtz (Routing reliability): Monitored Banyan during peak: fallback activation observed for runtime overload cases; overall 503 rate decreased, but gateway sheds still occur at extreme spikes. No evidence of fallback loops after hotfix. Marking resolved; tracking follow-up enhancement separately to support safe fallback on gateway breaker-open if Product approves.\nResolved with routing hotfix: loop-prevention/attempted_models bookkeeping bug prevented fallback from triggering in some overload scenarios. After deployment, fallback works when overload is detected at routing/runtime stages. Gateway admission shedding remains a terminal response by design (no fallback attempted) to prevent overload cascades; customer was informed and given mitigation guidance (Retry-After/backoff, optional policy/capacity adjustments).","textOnly":true},{"key":"SRC:0012","kind":"document","fileName":"Hosted API overload errors support guidance, retry backoff, and escalation.txt","score":0.8009990743549127,"isCited":true,"n":6,"localUrl":null,"liveUrl":null,"system":"Wiki","page":1,"fileId":"01a09931-3830-75a6-8a53-e40868aada53","text":"$3e","textOnly":true},{"key":"SRC:0013","kind":"document","fileName":"Hosted API overload errors support guidance, retry backoff, and escalation.txt","score":0.838774116135828,"isCited":true,"n":7,"localUrl":null,"liveUrl":null,"system":"Wiki","page":2,"fileId":"01a09931-3830-75a6-8a53-e40868aada53","text":"$3f","textOnly":true},{"key":"SRC:0009","kind":"document","fileName":"Retry-After header missing on some overload (503 429) responses for Hosted API (enterprise tenant).txt","score":0.7773121058661007,"isCited":true,"n":8,"localUrl":null,"liveUrl":null,"system":"Jira","page":1,"fileId":"01a09931-0ed6-7741-8867-c7342646f0ff","text":"$40","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Fallback experiment matrix cost penalty catalog (working notes).txt","score":0.7796722677881942,"isCited":false,"n":9,"localUrl":null,"liveUrl":null,"system":"Drive","page":0,"fileId":"01a09931-0ed3-730a-8f12-e2fffc6a5c5e","text":"$41","textOnly":true},{"key":"SRC:0002","kind":"document","fileName":"GA rollout checklist Hosted overload protection (circuit breakers load shedding).txt","score":0.7849667606677684,"isCited":false,"n":10,"localUrl":null,"liveUrl":null,"system":"Drive","page":0,"fileId":"01a09931-0ed3-730a-8f12-e30b6712c1f8","text":"$42","textOnly":true},{"key":"SRC:0008","kind":"document","fileName":"Retry-After header missing on some overload (503 429) responses for Hosted API (enterprise tenant).txt","score":0.8023253015620541,"isCited":false,"n":11,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c7342646f0ff","text":"$43","textOnly":true},{"key":"SRC:0010","kind":"document","fileName":"#all-hands - 2350009876 supply fallbacks model handoff clarify.txt","score":0.7783743398626224,"isCited":false,"n":12,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-236b-7265-918a-68c368146503","text":"$44","textOnly":true},{"key":"SRC:0011","kind":"document","fileName":"ADR-015 Admission control layering for hosted overload protection (Gateway vs Routing vs Runtime).txt","score":0.8218579408084761,"isCited":false,"n":13,"localUrl":null,"liveUrl":null,"system":"Wiki","page":2,"fileId":"01a09931-3830-75a6-8a53-e3c5f15aad28","text":"$45","textOnly":true},{"key":"SRC:0014","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.8900534361695397,"isCited":false,"n":1,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000000.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121109Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=3f33a29179536dd6ab5a6802204478e0e131bb8ff8e354e2a3df96630466cf2586748b783676fc054eff3ce5698e032db39b9e9039bb3a72c52aa0754396f018bc05feb55c0b58fa10a8e3813571bd99443b2bf60306f9be25d713996c22fe1da7be8887835493d9789163fb0d4b9016723f13c82bbf0b56a383b4f3c8e298f6f77adb9f7235e42b09ff58ffdb43abe478598d753a27165101216671256a6e10dacb8dcb69c97dc2381efe6ff27e9b93acd29e5df05c55bc43dc38011f97e24a45f19e99d030ce57c35d1e2aecfb109652f88b09e8f5282e61db7549f293dae82ee6181e757840f798b545707a631a986d12237ce5aabe560ffbd66bd5840653","system":"Recordings","start":0,"end":5,"transcript":"$46"},{"key":"SRC:0015","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.8539343103188929,"isCited":false,"n":2,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000004.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121109Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":20,"end":25,"transcript":"$47"},{"key":"SRC:0016","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.783768714261425,"isCited":false,"n":3,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000009.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121109Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":45,"end":50,"transcript":"$48"},{"key":"SRC:0017","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.783748327814019,"isCited":false,"n":4,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000011.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121109Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":55,"end":60,"transcript":"$49"},{"key":"SRC:0018","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.878097611576425,"isCited":false,"n":5,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000012.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121109Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":60,"end":65,"transcript":"$4a"},{"key":"SRC:0019","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.8229793953365542,"isCited":false,"n":6,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000013.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121109Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":65,"end":70,"transcript":"$4b"},{"key":"SRC:0020","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.784625949452016,"isCited":false,"n":7,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000014.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121109Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":70,"end":75,"transcript":"$4c"}],"secs":12.3,"effort":"standard","score":1472},"case-study-wording":{"answer":"$4d","sources":[{"key":"SRC:0003","kind":"document","fileName":"Northwind x Redwood public case study logo quote approval process.txt","score":0.9140072464942932,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed4-7228-bdd7-0426744e9410","text":"$4e","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Publish Northwind Commerce customer story standard metrics disclaimer.txt","score":0.9860824048519135,"isCited":true,"n":2,"localUrl":null,"liveUrl":null,"system":"GitHub","page":0,"fileId":"01a09931-0ed6-7741-8867-c6a1c6df810f","text":"$4f","textOnly":true},{"key":"SRC:0006","kind":"document","fileName":"Publish Northwind Commerce customer story standard metrics disclaimer.txt","score":0.7182526003544876,"isCited":true,"n":3,"localUrl":null,"liveUrl":null,"system":"GitHub","page":1,"fileId":"01a09931-0ed6-7741-8867-c6a1c6df810f","text":"$50","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Case study request permission to use Capitol Health logo & quote.txt","score":0.7116355615663252,"isCited":false,"n":4,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-02f50fbf61f1","text":"$51","textOnly":true},{"key":"SRC:0002","kind":"document","fileName":"Northwind x Redwood public case study logo quote approval process.txt","score":0.797598713659851,"isCited":false,"n":5,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-0426744e9410","text":"$52","textOnly":true},{"key":"SRC:0004","kind":"document","fileName":"Northwind x Redwood public case study logo quote approval process.txt","score":0.782699922219102,"isCited":false,"n":6,"localUrl":null,"liveUrl":null,"system":"Email","page":2,"fileId":"01a09931-0ed4-7228-bdd7-0426744e9410","text":"Subject: Re: Northwind x Redwood: public case study + logo/quote approval process\n\nMaya, Ben — great, thank you.\n\nNext steps on Redwood side:\n1) Marketing (Camila/Soojin) will deliver the full draft story + proposed quote text + draft diagram caption by Fri, Oct 17.\n2) We’ll keep metrics phrased as approximate ranges and clearly “observed” (no guarantees), and we’ll avoid any competitor references.\n3) Legal (Michael) will send the updated consent form via DocuSign incorporating Ben’s items (logo scope, approval rights, revocation timeline).\n\nTarget publish date: we’ll tentatively hold the week of Nov 3, and can pull earlier if approvals land sooner.\n\nThanks again — really appreciate Northwind partnering on this.\n\nBest,\nAvery Johnson\nEnterprise AE, Redwood Inference\n(415) 555-0184","textOnly":true},{"key":"SRC:0007","kind":"video","fileName":"Arm platform review - AmpereOne and Altra Max motherboards.mp4","score":0.7994901338738392,"isCited":false,"n":1,"localUrl":"/media/recordings/Arm platform review - AmpereOne and Altra Max motherboards.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-601fc4c98005/segments/segment_chunk_000000_000010.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":50,"end":55,"transcript":"$53"},{"key":"SRC:0008","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.7677631719613625,"isCited":false,"n":2,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000000.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":0,"end":5,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0009","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.7984916149370251,"isCited":false,"n":3,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000007.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":35,"end":40,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0010","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.8401172864142481,"isCited":false,"n":4,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000011.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=35b97880ecff1e986d1cbbc0ddd46d2ada98d8933ef6ec9fce59bf6cf4f777215e6200fd25187af46b103c25ccef975408ca4bd43f62b7ee630627d84105e32039e65ab2b6235cb29133b8e898911c227063d5265846e0f4c6baaf84009249597687752f589d746dfc09b3482792d4a5def26c83d31d47349be5851081767d783e221a3c48b2553b3f40cfd381f53ed65e916a4442b15930e90b98ca9764ea0b1aec69d1f61ebcfb301a992e4102042a3a69c936f54a2eea3dc3979d18777fbb0d203e7cd6ad809e786438aa6b75e9f5d64358d17ea9a0942daa09c9fb3a42bd69e03349f68f19dde33efd5a6c392aa14f6ed63c8e0b75d903710596c4d2715f","system":"Recordings","start":55,"end":60,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0011","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.7679384425180434,"isCited":false,"n":5,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000015.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":75,"end":80,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0012","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.7332116783081273,"isCited":false,"n":6,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000016.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":80,"end":85,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0013","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.7303003773771444,"isCited":false,"n":7,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000017.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":85,"end":90,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0014","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.7205885202896124,"isCited":false,"n":8,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000009.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":45,"end":50,"transcript":"$54"},{"key":"SRC:0015","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.7342859382125049,"isCited":false,"n":9,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000013.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":65,"end":70,"transcript":"$55"},{"key":"SRC:0016","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":0.7562924136020619,"isCited":false,"n":10,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000017.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":85,"end":90,"transcript":"$56"},{"key":"SRC:0017","kind":"video","fileName":"GPU serving node build - Epyc and Nvidia capacity testing.mp4","score":0.8401875563840978,"isCited":false,"n":11,"localUrl":"/media/recordings/GPU serving node build - Epyc and Nvidia capacity testing.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-602291f06671/segments/segment_chunk_000000_000016.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":80,"end":85,"transcript":"$57"},{"key":"SRC:0018","kind":"video","fileName":"GPU serving node build - Epyc and Nvidia capacity testing.mp4","score":0.810374880832833,"isCited":false,"n":12,"localUrl":"/media/recordings/GPU serving node build - Epyc and Nvidia capacity testing.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-602291f06671/segments/segment_chunk_000000_000017.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":85,"end":90,"transcript":"$58"},{"key":"SRC:0019","kind":"video","fileName":"Hardware evaluation - Supermicro Ampere and Grace Hopper serving nodes.mp4","score":0.7871045402381218,"isCited":false,"n":13,"localUrl":"/media/recordings/Hardware evaluation - Supermicro Ampere and Grace Hopper serving nodes.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-60238d69b94c/segments/segment_chunk_000000_000005.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":25,"end":30,"transcript":" scale I'm just gonna check this one out here in the front here it says these guys are having an event together with the grace grace hopper super chip that goes directly in this server right there so there you have it the latest super micro ARM servers so far it's not in a desktop But it's in these servers. Hey, what's up? How are you? And here is their server solutions. Maybe desktop's coming later? All right. So there it is with the top off. This is the Grace Hopper solution right here with Supermicro. And with all these, look at all these fast memory slots going around. PCIe and two big chips with a big cooling and all this connected with a grace hopper there it is look at those big power supplies"},{"key":"SRC:0020","kind":"video","fileName":"Hardware evaluation - Supermicro Ampere and Grace Hopper serving nodes.mp4","score":0.7333348765741566,"isCited":false,"n":14,"localUrl":"/media/recordings/Hardware evaluation - Supermicro Ampere and Grace Hopper serving nodes.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-60238d69b94c/segments/segment_chunk_000000_000007.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121133Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=0ebe0b85975a564530a97ed6ef4b80a99ffbe05d01fd4cc20fd9d3d8aa8ae21651948226605fd3a77dbb276b0ed5ee0fd703eb207df7dccb2cb24e12dfe540140b292906b4427bdfc3ee60e9f92d6415432475149b645c9afc0880b082d9e7b647f6d5e8667f75b08eeebc006e6d74c8f2f15512ef271df9d086c04895c83bb91a0d09c539f9c3ed0320a7c5e9ed12967133476ee98499906cfd93df822ddc72d50d43ed8ba2158f9e3b9351fb4138920adabc392235559ce17dabf30ea4295c080d9d6c8940144360f8ca9a79600a4356806c2bf8afd62c46da6d395b6cae96d3e2fbb8555e980a9c755bdad9b22251a457f90da89ff43be8b2100e0e61c84b","system":"Recordings","start":35,"end":40,"transcript":" scale I'm just gonna check this one out here in the front here it says these guys are having an event together with the grace grace hopper super chip that goes directly in this server right there so there you have it the latest super micro ARM servers so far it's not in a desktop But it's in these servers. Hey, what's up? How are you? And here is their server solutions. Maybe desktop's coming later? All right. So there it is with the top off. This is the Grace Hopper solution right here with Supermicro. And with all these, look at all these fast memory slots going around. PCIe and two big chips with a big cooling and all this connected with a grace hopper there it is look at those big power supplies"}],"secs":8.5,"effort":"standard","score":991},"merkle-anchoring":{"answer":"$59","sources":[{"key":"SRC:0004","kind":"document","fileName":"Admin activity chronicle and immutable-hash anchoring for audit evidence.txt","score":0.7849036945488813,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Linear","page":0,"fileId":"01a09931-0ed6-7741-8867-c780f72e3031","text":"$5a","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Audit ledgering and residency-aware evidence fabric for admin actions.txt","score":0.6637007968986774,"isCited":true,"n":2,"localUrl":null,"liveUrl":null,"system":"Linear","page":0,"fileId":"01a09931-0ed6-7741-8867-c78b382e0465","text":"$5b","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Fractured Context Session Anchoring & Context Interleaves - Notes.txt","score":0.5286041256023555,"isCited":false,"n":3,"localUrl":null,"liveUrl":null,"system":"Drive","page":0,"fileId":"01a09931-0ed3-730a-8f12-e30921d0c6da","text":"$5c","textOnly":true},{"key":"SRC:0002","kind":"document","fileName":"Fractured Context Session Anchoring & Context Interleaves - Notes.txt","score":0.7460618468486282,"isCited":false,"n":4,"localUrl":null,"liveUrl":null,"system":"Drive","page":1,"fileId":"01a09931-0ed3-730a-8f12-e30921d0c6da","text":"$5d","textOnly":true},{"key":"SRC:0003","kind":"document","fileName":"SUP 67219 append only signed audit stream and siem mapping request.txt","score":0.5501399682824804,"isCited":false,"n":5,"localUrl":null,"liveUrl":null,"system":"Jira","page":0,"fileId":"01a09931-0ed6-7741-8867-c751838ffb6e","text":"$5e","textOnly":true},{"key":"SRC:0006","kind":"document","fileName":"Tenant geo-guards retention enforcement and automated compliance exports for Dedicated.txt","score":0.4987441859825745,"isCited":false,"n":6,"localUrl":null,"liveUrl":null,"system":"Linear","page":0,"fileId":"01a09931-236b-7265-918a-682039e68b20","text":"$5f","textOnly":true},{"key":"SRC:0007","kind":"document","fileName":"Authorization Incident Containment Blueprint and Cross-Region Retention Procedure.txt","score":0.49564168103806694,"isCited":false,"n":7,"localUrl":null,"liveUrl":null,"system":"Wiki","page":1,"fileId":"01a09931-3830-75a6-8a53-e3cff6c008da","text":"$60","textOnly":true},{"key":"SRC:0008","kind":"video","fileName":"Arm platform review - AmpereOne and Altra Max motherboards.mp4","score":0.580211942436258,"isCited":false,"n":1,"localUrl":"/media/recordings/Arm platform review - AmpereOne and Altra Max motherboards.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-601fc4c98005/segments/segment_chunk_000000_000006.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121209Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=7c08acb338f5a786befa6b913401cf9a0701579b06cdfd31776fa891282f49f7e2de0f47f98c5dec466a51368e9e6247096141e2860e21fe1859f6550f4e0ce1acdb90b5025ea91c4dafda621e6d86587cdf9424faab7cf5c957797aca94a1d0f23776b857e2ec77ec1a866e0e253261041bf25983b0d13c7ecc817de448768ccce70c811fc04e47532e39fa308bb26d1ba1e86a85186a60079e7a969d84540f84aed4c0846cdeee735512636a7ee6af19f9ec6e4289c84c17bdc1b2f2e19f1bc6faaa994e9559841cc451b436f57111aec0f10db5f5ca7a0485eea213ce7d232413c23b0e85149dbde212aa816c023ff8af53290ee2a1bafdb1d7a782925832","system":"Recordings","start":30,"end":35,"transcript":"$61"}],"secs":8.6,"effort":"standard","score":564},"mission":{"answer":"Redwood Inference’s mission statement is: **“Helping teams ship consistent, cost-effective inference.”** [[SRC:0009]]","sources":[{"key":"SRC:0009","kind":"document","fileName":"Intro demo request to pivot to technical workshop.txt","score":1.0273597240447998,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Email","page":2,"fileId":"01a09931-0ed4-7228-bdd7-03c40342f01c","text":"\"Helping teams ship consistent, cost-effective inference.\"\n\"redwood.com\"\n\"\"\n","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Aligning infra owners slot for technical deep-dive.txt","score":0.7389344899704542,"isCited":false,"n":2,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-02b699311302","text":"$62","textOnly":true},{"key":"SRC:0002","kind":"document","fileName":"Aligning infra owners slot for technical deep-dive.txt","score":0.7299504516863553,"isCited":false,"n":3,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed4-7228-bdd7-02b699311302","text":"$63","textOnly":true},{"key":"SRC:0003","kind":"document","fileName":"Clarifying which SOC 2 artifact is authoritative (bridge reissue status).txt","score":0.9172648032652635,"isCited":false,"n":4,"localUrl":null,"liveUrl":null,"system":"Email","page":2,"fileId":"01a09931-0ed4-7228-bdd7-0305e821c63f","text":"Trust\\nRedwood Inference\\n","textOnly":true},{"key":"SRC:0004","kind":"document","fileName":"Exec intro technical deep-dive planning availability & next steps.txt","score":0.7328879356040731,"isCited":false,"n":5,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-037669432851","text":"$64","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Integrated Launch Assurance GTM, Enablement, Pricing & Support Pre-QBR Checklist.txt","score":0.8734017304377542,"isCited":false,"n":6,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-03b8ebe1a8b0","text":"$65","textOnly":true},{"key":"SRC:0006","kind":"document","fileName":"Integrated Launch Assurance GTM, Enablement, Pricing & Support Pre-QBR Checklist.txt","score":1.0292858183383942,"isCited":false,"n":7,"localUrl":null,"liveUrl":null,"system":"Email","page":2,"fileId":"01a09931-0ed4-7228-bdd7-03b8ebe1a8b0","text":"Redwood Inference\n","textOnly":true},{"key":"SRC:0007","kind":"document","fileName":"Intro Oliver Park Senior ML Infrastructure (interview loop slots).txt","score":0.7300086473725445,"isCited":false,"n":8,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed4-7228-bdd7-03c0bcb9d1ff","text":"$66","textOnly":true},{"key":"SRC:0008","kind":"document","fileName":"Intro demo request to pivot to technical workshop.txt","score":0.9065368861859653,"isCited":false,"n":9,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-03c40342f01c","text":"$67","textOnly":true},{"key":"SRC:0010","kind":"document","fileName":"MediSphere onboarding steward assignment & kickoff plan.txt","score":0.7338436711040508,"isCited":false,"n":10,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed4-7228-bdd7-040520285ca6","text":"$68","textOnly":true},{"key":"SRC:0011","kind":"document","fileName":"Ops ML schedule 60 90m technical deep-dive (timezones & attendee roster).txt","score":0.7783134594089244,"isCited":false,"n":11,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed5-7746-ba56-63e04059dd72","text":"$69","textOnly":true},{"key":"SRC:0012","kind":"document","fileName":"Stakeholder scheduling multi-session demo options after travel & procurement blockers.txt","score":0.7640240735225545,"isCited":false,"n":12,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed5-7746-ba56-650423d84d08","text":"$6a","textOnly":true},{"key":"SRC:0013","kind":"document","fileName":"Stakeholder scheduling multi-session demo options after travel & procurement blockers.txt","score":0.743594740849998,"isCited":false,"n":13,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed5-7746-ba56-650423d84d08","text":"$6b","textOnly":true},{"key":"SRC:0014","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.8334760779351179,"isCited":false,"n":1,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000007.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121221Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":35,"end":40,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0015","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.7368645250438268,"isCited":false,"n":2,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000010.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121221Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":50,"end":55,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0016","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.8769255027697014,"isCited":false,"n":3,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000011.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121221Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":55,"end":60,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0017","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.8015842736382667,"isCited":false,"n":4,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000015.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121221Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":75,"end":80,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0018","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.7653360191899826,"isCited":false,"n":5,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000016.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121221Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":80,"end":85,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0019","kind":"video","fileName":"Arm server evaluation - 128-core Altra Max for cost per token.mp4","score":0.762297164884874,"isCited":false,"n":6,"localUrl":"/media/recordings/Arm server evaluation - 128-core Altra Max for cost per token.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6020c1582296/segments/segment_chunk_000000_000017.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121221Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":85,"end":90,"transcript":" Is it possible I can buy one? Sure. It's for everybody? Yeah, it's everybody. And does it fit in a box like this? It's going to fit something like this? Well, we need to fit, because this is our first board in our Paralyze, so all of our team is now designing the related chassis for now. So that could be server chassis, could be desktop chassis, we don't know. We will be the server chassis. Server chassis? Yes. I see four potential PCIe slots, right? Yes. So it could be GPU? Yes. It can support up to two double-width GPUs. And a lot of RAM. Yeah, it can support up to eight RAMs in our micro edX motherboard. And it can support... the previous and the next ampere maybe it will be current ultra max and ultra all right uh can you show a little bit what you show here at the show so what do you have you have a lot of servers workstations yes so asrock rack is specialized making server and workstation motherboard motherboard and bare bones so for from here we are starting from the intel and so this intel yeah"},{"key":"SRC:0020","kind":"video","fileName":"Hardware evaluation - Supermicro Ampere and Grace Hopper serving nodes.mp4","score":0.7654646151647652,"isCited":false,"n":7,"localUrl":"/media/recordings/Hardware evaluation - Supermicro Ampere and Grace Hopper serving nodes.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-60238d69b94c/segments/segment_chunk_000000_000007.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260915%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260915T121221Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":35,"end":40,"transcript":" scale I'm just gonna check this one out here in the front here it says these guys are having an event together with the grace grace hopper super chip that goes directly in this server right there so there you have it the latest super micro ARM servers so far it's not in a desktop But it's in these servers. Hey, what's up? How are you? And here is their server solutions. Maybe desktop's coming later? All right. So there it is with the top off. This is the Grace Hopper solution right here with Supermicro. And with all these, look at all these fast memory slots going around. PCIe and two big chips with a big cooling and all this connected with a grace hopper there it is look at those big power supplies"}],"secs":3.8,"effort":"standard","score":558},"hardware-capacity":{"answer":"$6c","sources":[{"key":"SRC:0001","kind":"document","fileName":"Regional capacity forecast & rollout timeline Q4-2025 to Q2-2026.txt","score":1,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Drive","page":0,"fileId":"01a09931-0ed3-730a-8f12-e35340da5487","text":"$6d","textOnly":true},{"key":"SRC:0003","kind":"image","fileName":"GPU capacity and cost by region.jpg","score":1,"isCited":true,"n":1,"localUrl":"/media/diagrams/GPU capacity and cost by region.png","liveUrl":"https://storage.googleapis.com/graphon-private-upload-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0993f-ebed-774e-a017-b978ecae4760/GPU%20capacity%20and%20cost%20by%20region.jpg?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260916%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260916T033054Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Diagrams"},{"key":"SRC:0002","kind":"document","fileName":"#eng-platform - 1960012345 cross region egress and commitment breakdown.txt","score":1,"isCited":true,"n":2,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-236d-7700-af57-aad76a825052","text":"$6e","textOnly":true},{"key":"SRC:0004","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":1,"isCited":true,"n":1,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000001.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260916%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260916T033054Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":5,"end":10,"transcript":"$6f"},{"key":"SRC:0008","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":1,"isCited":true,"n":2,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000005.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260916%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260916T033054Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":25,"end":30,"transcript":"$70"},{"key":"SRC:0005","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":1,"isCited":true,"n":3,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000002.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260916%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260916T033054Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":10,"end":15,"transcript":"$71"},{"key":"SRC:0006","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":1,"isCited":true,"n":4,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000003.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260916%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260916T033054Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":15,"end":20,"transcript":"$72"},{"key":"SRC:0007","kind":"video","fileName":"Datacenter walkthrough - decommissioned facility, racks and power.mp4","score":1,"isCited":true,"n":5,"localUrl":"/media/recordings/Datacenter walkthrough - decommissioned facility, racks and power.mp4","liveUrl":"https://storage.googleapis.com/graphon-private-processed-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0a4e0-1b28-7005-a3fb-6021f9691fac/segments/segment_chunk_000000_000004.mp4?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260916%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260916T033054Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Recordings","start":20,"end":25,"transcript":"$73"}],"secs":54.1,"effort":"ultra","score":2100},"gpu-card":{"answer":"$74","sources":[{"key":"SRC:0006","kind":"document","fileName":"#eng-infra - 1713571200 new dedicated gpu launch paris mumbai.txt","score":1,"isCited":true,"n":1,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-236c-76c2-a851-4b40b10ff485","text":"$75","textOnly":true},{"key":"SRC:0003","kind":"document","fileName":"Dedicated capacity recommendation & ops playbook for Sparkle Inc H100 A100 L4 mix.txt","score":1,"isCited":true,"n":2,"localUrl":null,"liveUrl":null,"system":"Email","page":0,"fileId":"01a09931-0ed4-7228-bdd7-034648cafe83","text":"$76","textOnly":true},{"key":"SRC:0007","kind":"document","fileName":"#eng-infra - 1968001111 vendor capacity handshake.txt","score":1,"isCited":true,"n":3,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-236c-76c2-a851-4b7a205a01c2","text":"$77","textOnly":true},{"key":"SRC:0008","kind":"document","fileName":"#eng-infra - 7365432100 apac h100 canary ops billing clarify.txt","score":1,"isCited":true,"n":4,"localUrl":null,"liveUrl":null,"system":"Slack","page":0,"fileId":"01a09931-236c-76c2-a851-4b902e4c0067","text":"$78","textOnly":true},{"key":"SRC:0001","kind":"document","fileName":"Regional capacity forecast & rollout timeline Q4-2025 to Q2-2026.txt","score":1,"isCited":true,"n":5,"localUrl":null,"liveUrl":null,"system":"Drive","page":0,"fileId":"01a09931-0ed3-730a-8f12-e35340da5487","text":"$79","textOnly":true},{"key":"SRC:0009","kind":"image","fileName":"GPU capacity and cost by region.jpg","score":1,"isCited":true,"n":1,"localUrl":"/media/diagrams/GPU capacity and cost by region.png","liveUrl":"https://storage.googleapis.com/graphon-private-upload-g000000001/users/user_01KZC4Q9J4E0RGJXY2NGT8PX5M/files/01a0993f-ebed-774e-a017-b978ecae4760/GPU%20capacity%20and%20cost%20by%20region.jpg?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=sa-compute-engine%40graphon-ai-g000000001.iam.gserviceaccount.com%2F20260921%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20260921T125800Z&X-Goog-Expires=900&X-Goog-SignedHeaders=host&X-Goog-Signature=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","system":"Diagrams"},{"key":"SRC:0004","kind":"document","fileName":"Dedicated reservation questionnaire and fallback plan.txt","score":1,"isCited":true,"n":6,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed4-7228-bdd7-034a7d8f5e83","text":"$7a","textOnly":true},{"key":"SRC:0002","kind":"document","fileName":"Cluster sizing & scaling path for Saturn Health dedicated capacity.txt","score":1,"isCited":false,"n":7,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed4-7228-bdd7-030e7f0e3246","text":"$7b","textOnly":true},{"key":"SRC:0005","kind":"document","fileName":"Scheduling GPU capacity architecture review scaling & isolation (calendar update).txt","score":1,"isCited":false,"n":8,"localUrl":null,"liveUrl":null,"system":"Email","page":1,"fileId":"01a09931-0ed5-7746-ba56-64ef538d9c52","text":"$7c","textOnly":true}],"secs":68.7,"effort":"ultra","score":1975}}}],["$L7d","$L7e"],"$L7f"]}],{},null,false,null]},null,false,null],"$L80",false]],"m":"$undefined","G":["$81",["$L82"]],"S":true,"h":null,"r":"$undefined","s":"$undefined","a":"$undefined","l":"$undefined","p":"$undefined","d":"$undefined"}

👋 You’re looking at #capacity