---
type: briefing
category: daily
domain: ai-infrastructure
status: final
captured: 2026-05-10
updated: 2026-05-10
confidence: medium
tags:
  - briefing
  - ai-infrastructure
  - ai-factory
  - servicenow
  - iren
  - mi350p
  - inference-economics
  - power
  - intel-foundry
note_role: briefing
canonical: false
---
# 2026-05-10 AI Infrastructure Brief

## Executive read

Two weekend-shaped updates on top of the prior systemic read. **The "AI factory" stopped being a hardware story and started becoming a distribution-and-governance bundle:** NVIDIA's Knowledge26 keynote with ServiceNow rolled out **Project Arc + NVIDIA OpenShell + Vibe Coding**, with Cisco doing the silicon-to-software security wrap, Microsoft Agent 365 carrying enterprise governance, and Anthropic's Claude as the default model for Building. In parallel, NVIDIA backed **IREN's reported 5 GW DSX-architecture pipeline** ($3.4B cloud contract + up to ~$2.1B equity stake; 2 GW Sweetwater Texas as the reference site) — a clean neocloud-to-hyperscale bridge that signals where named capacity goes when hyperscaler queues are stuck on power. Two reliability nudges sharpen the read: **AWS US-EAST-1 had a Northern Virginia data-center cooling failure on May 8** that took Coinbase dark for ~7 hours, and **Constellation/Microsoft confirmed the Three Mile Island restart** as the Crane Clean Energy Center for 2027 ($1.6B, 20-year PPA, 835 MW carbon-free). Power and concentration risk remain the marginal story; what changed this weekend is the **packaging layer above the chip**.

## Top infrastructure signals

1. **NVIDIA + ServiceNow Knowledge26 turns the AI Factory into an enterprise control plane.**
   - Jensen Huang and ServiceNow CEO Bill McDermott unveiled **Project Arc** and **Vibe Coding**, integrating the NVIDIA AI Factory inside the ServiceNow platform. `@nvidia` framed it as "the next chapter of enterprise AI."
   - Multiple independent posters describe the shape as **Project Arc = secure desktop/agent surface inside ServiceNow, running on NVIDIA OpenShell sandboxes**, with ServiceNow's AI Control providing governance.
   - The full vendor stack visible at the event: **NVIDIA (compute + OpenShell sandbox), Cisco (Secure AI Factory: switching/security), ServiceNow (workflow/governance), Microsoft (Agent 365 governance integration), Anthropic (Claude as the default model for the "Building" surface)**. `@m_buvik` summarized this as ServiceNow becoming "deeply locked in" with all four.
   - **Implication:** the AI Factory is now sold less as a rack and more as an **enterprise distribution + governance bundle**. The differentiator moves from GPU SKU to *which workflow plane* the factory plugs into. That changes the buying conversation from "how many H100/B200/Rubin per cluster" to "which control plane do my agents live inside, and who governs them."
   - **Operator caveat (jonashernlund, May 9):** "AI Factory output is gated by ServiceNow CMDB hygiene, not model quality. Enterprises with messy asset data discover this in week 3." This is the durable critique — the integration story is the easy part; cleaning years of stale CMDB records is the actual project.

2. **NVIDIA × IREN: reported 5 GW DSX pipeline, $3.4B cloud + ~$2.1B equity, 2 GW Sweetwater Texas reference site.**
   - Posted on May 8 (`@TeckNexus`, `@XIRRAdvisors`, others; per Data Center Knowledge): NVIDIA backing IREN's 5 GW AI factory pipeline using the DSX architecture; $3.4B cloud contract, potential ~$2.1B equity stake. The 2 GW Sweetwater (TX) campus is positioned as the reference deployment.
   - Treat the dollar figures as reported / pending IREN's own filing confirmation; the directional commitment and DSX-as-template signal is the core read regardless of final numbers.
   - **Implication:** when hyperscaler queues are gridlocked on power, **NVIDIA-backed neocloud capacity becomes a routable substitute** for named compute demand. DSX as a "reference architecture" frames how subsequent NVIDIA-aligned capacity (Crusoe, CoreWeave, Nebius, others) gets standardized. This also widens the set of operators that can credibly bid on sovereign-AI capacity in non-US geographies.

3. **AMD Instinct MI350P PCIe: drop-in air-cooled inference is real.**
   - `@AMD` (May 8): "AMD Instinct MI350P PCIe GPUs are designed to deliver production AI inference performance in **standard air-cooled servers**, helping enterprises scale AI with existing infrastructure."
   - **Implication:** counterbalance to the all-rack-scale-liquid-cooled narrative. Most existing colos and enterprise data halls are air-cooled, with installed PDU budgets that don't fit Rubin-class densities. A credible PCIe form factor at MI350-class capability moves AMD's TAM from "new hyperscaler greenfield" into **installed-base inference modernization** — exactly the EMEA/Lenovo hybrid story. Pair this with AMD's Pollara/UEC posture and the fabric-choice story, and AMD now has a coherent two-lane offering: rack-scale UEC for new builds, PCIe drop-in for retrofit.

4. **Single-zone reliability is back on the table: AWS US-EAST-1 cooling failure, ~7-hour Coinbase outage (May 8).**
   - Per multiple X reports, an AWS data-center cooling failure in Northern Virginia took US-EAST-1 down; Coinbase reported ~7 hours of trading downtime. AWS rerouted traffic; some hardware was reportedly damaged.
   - **Implication:** as workloads concentrate in a few power-rich locations, **single-zone exposure is now a top-of-stack risk** — not just a database/availability concern. Operators with multi-zone, multi-region (or hybrid on-prem + cloud) AI inference paths price reliability differently. The buyer question after this weekend: *"How does my agentic workflow degrade if my primary region goes dark for 7 hours?"* For sovereign / regulated EMEA workloads, this strengthens the multi-site mandate rather than weakens it.

5. **DeepSeek V4 inference on GB300: 4× iso-interactivity throughput improvement.**
   - SemiAnalysis (May 9): work by `@sgl_project` and `@radixark` on DeepSeek V4 inference on B200/B300 yielded "the recent 4× iso-interactivity throughput improvements on GB300" by `@ChengWan17`. Quoted Elon: *"the GB300 is the b—."*
   - In parallel, `@NVIDIAAI` (May 8): an ICML26 paper with `@SakanaAILabs` on **TwELL sparse packing + fused CUDA kernels** delivered "20%+ inference/training speedups at scale."
   - **Implication:** inference economics is still moving on the silicon people already have. **A meaningful share of 2026 cost-down comes from software/format work on existing GB200/GB300/B200 fleets, not just from new silicon launches.** This partially decouples cost-per-token trajectory from new-capacity build pace, which matters when ~30–50% of planned 2026 capacity is reportedly slipping to 2028. Buyers should ask their providers what their token-cost roadmap looks like *without* assuming Rubin-class capacity is online.

6. **Three Mile Island restart confirmed for 2027 — Crane Clean Energy Center, $1.6B Constellation + 20-year Microsoft 835 MW PPA.**
   - Reported broadly (May 10, e.g. `@Tickerwire`): Constellation targeting 2027 restart, $1.6B investment, 20-year Microsoft PPA, 835 MW carbon-free.
   - **Implication:** long-duration power procurement is now table-stakes for hyperscaler AI capacity. Pairs with prior week signals (Hut 8 15-year, 352 MW Beacon Point lease at $9.8B base / $25.1B max; Anthropic ↔ SpaceX 300+ MW Colossus 1 access; Stargate UAE 1 GW with first 200 MW live in 2026; Microsoft + G42 additional 200 MW UAE). The procurement primitive has clearly shifted from spot/cloud to **multi-decade offtake-style contracts often with embedded financing**. EU sovereign-AI buyers without comparable nuclear/baseload PPAs in flight will quickly find the gap material.

## Market structure / economics

- **AI factory = bundle, not a product.** The structural beneficiary of the Knowledge26 stack is the *workflow surface vendor* (ServiceNow today; potentially Salesforce, Workday, SAP, Oracle Fusion, HubSpot tomorrow). NVIDIA captures the compute and the sandbox layer. The integration tax — connectors, governance, identity — accrues to whichever vendor anchors the surface. For pure GPU resellers, the margin compression risk is real if "where does the agent run?" beats "what GPU is in the rack?"
- **Software-side inference gains are partially offsetting capacity slip.** GB300 4× iso-interactivity + sparse-format kernel speedups + Dynamo harness reuse improvements compound. This means **deployable token-throughput per MW is still climbing on existing fleets**, even as new build slips. Buyers with sticky workloads (especially agentic, where KV cache reuse is high-value) get a tailwind that capacity headlines understate.
- **Reliability is becoming a differentiator, not a baseline.** US-EAST-1 cooling outage maps to AI-inference availability at exactly the moment when more enterprise workflows go agentic and start requiring sub-second tool-call response. Multi-region inference routing (Cloudflare-style, but for AI providers) becomes a real product category.
- **Foundry geopolitics: Patrick Moorhead spent the weekend on the Apple ↔ Intel Foundry pressure story** — WSJ reporting it; Moorhead notes a likely USG $10B Intel investment, Apple as MFN customer, and NVIDIA likely the next high-performance MFN customer. The Intel Foundry/Design split commentary continues. **Implication for AI infra:** chip supply diversity is being state-engineered; expect this to feed back into which sovereign-AI factory builds get political support in 2H 2026.

## Enterprise / EMEA / Lenovo angle

- **The buying conversation just shifted.** "Which AI factory?" now means choosing a **bundle** (NVIDIA + ServiceNow + Cisco + Microsoft Agent 365 + Anthropic, or alternative compositions). Lenovo's hybrid-AI-factory thesis remains intact, but the articulation must shift from "we integrate the rack" to "we operationalize *this control plane* with sovereign and audit constraints **and** these workflow surfaces." Without an explicit ServiceNow / Microsoft / Salesforce-class workflow story, "system integrator" risks losing the buyer's attention to "control-plane vendor."
- **MI350P PCIe is a real Lenovo opportunity.** It maps cleanly to EMEA installed-base air-cooled racks, including a lot of mid-market enterprise and many EuroHPC sites. Drop-in inference modernization, governed by Lenovo's hybrid posture, with AMD silicon, is a coherent EMEA-friendly offer that doesn't require greenfield power+cooling buildouts.
- **Multi-zone resilience is now a procurement requirement.** AWS US-EAST-1 + Coinbase teaches the lesson at scale. EMEA-regulated buyers (financial services, public sector, healthcare) should treat single-region AI inference as a known operational gap. Lenovo's hybrid + on-prem story plus governed multi-site cloud routing reads well into this — the EMEA equivalent of "private subnet for primary inference, public cloud for burst" is a legitimate procurement frame.
- **Power procurement is decade-scale.** Constellation/MSFT 20-year 835 MW PPA, Hut 8 15-year, Anthropic-SpaceX, Stargate UAE — every named flagship is now multi-decade. EMEA sovereign-AI projects (France 1 GW Fluidstack, EuroHPC follow-ons, UAE-aligned EU partnerships) need to articulate equivalent baseload commitments or accept dependency on Gulf gigawatt clusters. This is a real political conversation, not a procurement detail.
- **NVIDIA × IREN is a template to watch.** If DSX-architecture neocloud capacity scales as a NVIDIA-backed substitute, EMEA could see a similar template land via Lenovo + a regional neocloud partner — as a structured response to slow EU sovereign-AI buildout. Worth tracking which European operators get DSX-equivalent backing.

## What to watch next

- **Knowledge26 follow-through:** real deployment data and case studies from ServiceNow Project Arc + NVIDIA OpenShell. The CMDB-hygiene caveat is the leading indicator of whether this becomes durable enterprise fact or another cycle of integrator slideware.
- **IREN deal confirmation and DSX template propagation.** Watch IREN filings for the actual contract structure, and watch which neoclouds get the next NVIDIA-backed DSX commit.
- **Constellation/MSFT replication.** Whether AWS, Google, Meta, or Oracle announce comparable nuclear / long-duration PPAs in the next 4–8 weeks. If yes, it's a sector-wide procurement re-rate. If no, Microsoft has a structural advantage on power supply that shows up in 2027–2028 deployable capacity.

## Noise to discount

- **Generic "AI infra" capex headline reposts** without earnings-call/filing support. Treat the ~$725B Big Four 2026 base as the durable anchor; everything above $800B is analyst variance.
- **"Stargate" as an X search term is heavily polluted** by sci-fi/fan content (Stargate Atlantis frames, Kurt Russell takes, $STG token spam). Filter on context, not just the keyword.
- **Decentralized GPU marketplace pitches** ("Cocoon," other dGPU aggregators) — tracked, but few have credible production-scale token-economics or audit/governance stories. Not a near-term sovereign-AI substitute.
- **"MFN customer" speculation about Intel Foundry without contract terms.** Moorhead's signal is directionally useful as analyst framing; treat the named-customer specifics as not-yet-priced until contracts are filed.

## Source trail

### Primary / verification sources

- NVIDIA × ServiceNow Knowledge26 announcement (NVIDIA on X): https://x.com/nvidia/status/1923163847... (`@nvidia`, 2026-05-08, 23:23 UTC) — Project Arc, Vibe Coding, NVIDIA AI Factory inside ServiceNow.
- AMD Instinct MI350P PCIe announcement (AMD on X): https://x.com/AMD (`@AMD`, 2026-05-08, 15:00 UTC) — production AI inference in standard air-cooled servers.
- NVIDIA × IREN reportage (DCK / multiple posters): https://x.com/TeckNexus, https://x.com/XIRRAdvisors (May 8) — 5 GW pipeline, DSX, Sweetwater TX.
- AWS US-EAST-1 cooling failure / Coinbase outage: multiple May 10 X posts (`@BSCNews`, `@WOLF_Crypto_X`); awaiting AWS post-mortem.
- Constellation Energy + Microsoft 835 MW Three Mile Island restart (Crane Clean Energy Center): `@Tickerwire`, May 10; Constellation IR for primary.
- Anthropic "Teaching Claude why" (cross-relevant, fuller treatment in Agentic AI Brief): https://www.anthropic.com/research/teaching-claude-why
- NVIDIA Dynamo harness path (NVIDIA AI on X): `@NVIDIAAI`, 2026-05-08, 19:36 UTC — Dynamo improvements for Claude Code, OpenClaw harnesses (cross-relevant).
- TwELL sparse + fused CUDA paper (NVIDIA × Sakana AI Labs, ICML 2026): `@NVIDIAAI`, 2026-05-08, 17:24 UTC.

### X discovery layer

- **`@SemiAnalysis_` (2026-05-09 01:00 UTC, ~257 lk):** *"Amazing work from the @sgl_project and @radixark team for their work optimizing DeepSeek V4 inference on B200, B300, and the recent 4× iso-interactivity throughput improvements on GB300 by @ChengWan17!"* — durable inference-economics signal: software-side gains continue on existing silicon.
- **`@SemiAnalysis_` (2026-05-09 17:00 UTC, ~291 lk):** *"In WW3, wars will be won based on the # of tokens that each country has."* — tokens-as-strategic-resource framing; reinforces sovereign-AI capacity story.
- **`@nvidia` (2026-05-08 23:23 UTC, ~448 lk):** Knowledge26 keynote Jensen × McDermott — Project Arc + Vibe Coding + AI Factory inside ServiceNow.
- **`@NVIDIAAI` (2026-05-08 19:36 UTC, ~318 lk):** Dynamo harness-facing path improvements for Claude Code, OpenClaw — KV cache reuse, tool parsing, cold-tool firing fixes.
- **`@NVIDIAAI` (2026-05-08 17:24 UTC, ~501 lk):** ICML26 collaboration with Sakana AI Labs — TwELL sparse packing + fused CUDA = 20%+ inference/training speedups at scale.
- **`@AMD` (2026-05-08 15:00 UTC, ~303 lk):** MI350P PCIe — production AI inference in standard air-cooled servers.
- **`@TeckNexus` / `@XIRRAdvisors` (2026-05-08):** NVIDIA × IREN 5 GW DSX pipeline, $3.4B cloud + ~$2.1B equity, 2 GW Sweetwater Texas.
- **`@PatrickMoorhead` weekend thread (May 8–9):** Intel Foundry / Apple WSJ pressure story; USG $10B Intel investment; NVIDIA likely next MFN high-performance customer; Intel Foundry/Design split commentary.
- **`@jonashernlund` (May 9, low likes but high signal):** AI Factory output gated by ServiceNow CMDB hygiene — operator caveat for Project Arc.
- **`@m_buvik` (May 8):** $NOW lock-in across NVIDIA (Project Arc + OpenShell + AI factories), Microsoft (Agent 365 governance), Anthropic (Claude as default Building model).
- **`@Gruve_ai` / `@Supermicro` (May 8):** Cisco + NVIDIA Secure AI Factory, full-stack silicon-to-software security; Supermicro + DDN + NVIDIA "AI Factory on Wheels."
- **`@Tickerwire` (May 10):** Three Mile Island Crane Clean Energy Center 2027 restart, $1.6B Constellation + 20-yr 835 MW Microsoft PPA.
- **`@BSCNews` / `@WOLF_Crypto_X` (May 10):** AWS US-EAST-1 cooling failure, ~7h Coinbase outage.

## KB updates completed

This brief was produced via a manual exception run (Opus 4.7) after the scheduled `Noam AI Infrastructure Morning Brief` job failed at 06:30 Asia/Jerusalem due to OpenAI quota exhaustion (FailoverError + openai-codex provider cooldown). To keep scope contained, **no durable KB note edits were performed in this run**. The following are flagged as durable update candidates for the next einstein cycle:

- `[[03 Research/AI Infrastructure/AI Infrastructure Overview]]` — promote the "AI Factory as control plane bundle" framing; add Knowledge26 / Project Arc / OpenShell + Cisco + Microsoft Agent 365 + Anthropic stack composition.
- `[[04 Companies/NVIDIA]]` — add Knowledge26 ServiceNow integration + IREN 5 GW DSX backing as durable enterprise-distribution and neocloud-anchor signals.
- `[[04 Companies/AMD]]` — add MI350P PCIe drop-in air-cooled inference as installed-base modernization vector (pair with Pollara/UEC for the two-lane offering).
- `[[04 Companies/ServiceNow]]` — create or expand: ServiceNow as multi-vendor agent control plane (NVIDIA + Microsoft + Anthropic), with the CMDB-hygiene operator caveat.
- `[[03 Research/AI Infrastructure/Inference Economics]]` — log GB300 4× iso-interactivity (sgl_project / radixark / ChengWan17) and TwELL sparse + fused CUDA (NVIDIA × Sakana, ICML26) as software-side cost-down evidence.
- `[[03 Research/AI Infrastructure/Sovereign AI]]` — add Constellation/MSFT 835 MW Three Mile Island PPA; flag IREN/DSX as a NVIDIA-backed neocloud template that EU/EMEA sovereign-AI projects could mirror.
- `[[05 People/Patrick Moorhead]]` — refresh with weekend Intel Foundry geopolitics framing; pair with primary WSJ-cited reporting when available.
- `[[05 People/Dylan Patel]]` — refresh with the May 9 SemiAnalysis posts (DeepSeek V4 GB300 throughput; tokens-as-strategic-resource).

## Executive Telegram version

*(Telegram delivery skipped on user instruction for this manual exception run.)*

For convenience, a short executive read:

AI Infra signal for May 10: **the AI factory just turned into a control-plane bundle, not a rack.**

- **NVIDIA × ServiceNow Knowledge26: Project Arc + OpenShell + Vibe Coding** — AI Factory inside ServiceNow, with **Cisco** for security, **Microsoft Agent 365** for governance, **Anthropic Claude** as default model for "Building." Buying conversation shifts from "which GPU?" to "which workflow plane runs my agents?" Operator caveat: ServiceNow CMDB hygiene gates output (jonashernlund).
- **NVIDIA × IREN reported 5 GW DSX pipeline** — $3.4B cloud + ~$2.1B equity, 2 GW Sweetwater Texas as reference. Neocloud-to-hyperscale bridge for power-blocked demand.
- **AMD MI350P PCIe (drop-in air-cooled inference)** — credible installed-base modernization path; counterbalance to liquid-cooled rack-only narrative. Real Lenovo + EMEA opportunity.
- **AWS US-EAST-1 cooling failure / Coinbase 7-hr outage (May 8)** — single-zone reliability is now a top-of-stack risk for agentic workflows. Multi-region inference routing becomes a real product category.
- **Constellation/MSFT 835 MW Three Mile Island restart confirmed for 2027** ($1.6B, 20-yr PPA). Long-duration power procurement is table-stakes; EU sovereign-AI without comparable PPAs has a visible gap.
- **Inference economics still moving on the silicon people already have:** SemiAnalysis on DeepSeek V4 GB300 4× iso-interactivity (sgl_project / radixark); NVIDIA × Sakana TwELL sparse + fused CUDA = 20%+ speedup. Software-side cost-down partially offsets reported 30–50% capacity slip.
- **Patrick Moorhead weekend thread on Intel Foundry geopolitics** (USG $10B, Apple MFN, NVIDIA likely next) — chip-supply diversity is being state-engineered; feeds into 2H sovereign-AI political support.

What matters for Lenovo / EMEA: differentiation now lives at the **workflow plane** layer above the rack, not the rack itself. Pair the hybrid-AI-factory story explicitly with a control-plane choice (ServiceNow / Microsoft / Salesforce-class) and articulate sovereign-AI procurement around multi-zone reliability + decade-scale power PPAs. MI350P PCIe is a real installed-base play. NVIDIA × IREN is the template to watch for an equivalent EMEA neocloud-anchor structure.

Sources: NVIDIA × ServiceNow X post, AMD MI350P X post, IREN/NVIDIA DCK reportage, AWS US-EAST-1 reports, Constellation/MSFT Three Mile Island, SemiAnalysis on DeepSeek V4 GB300, NVIDIA × Sakana ICML26 paper, Anthropic alignment research (cross-domain).
