brianletort.ai
All issues

The AI Stack Weekly

Issue 05 · Week 21 of 2026.

/Industry brief · ~7 min read/Public sources onlyDownload brief

The Bottom Line

The bull thesis got every confirmation it needed. The strategic battleground moved to runtimes.

Flywheel arcAll three lenses

W21 was the loudest week of Q2 across every lens at once. NVIDIA Q1 FY27 (Tuesday May 20) printed $81.6B revenue (+85% YoY) with Data Center at $75.2B (+92% YoY, +21% QoQ from $62.3B) and a Q2 guide of $91B with zero China DC compute assumed; total supply commitments stepped to $145B; $80B additional buyback authorized and the dividend lifted 25x. Huang said NVIDIA 'will be supply-constrained through the entire life of Vera Rubin' (Q3 2026 production) — the bear thesis that hyperscaler capex isn't translating to chip revenue is gone, and the binding constraint moved firmly to HBM4 and CoWoS packaging.

The frontier-lab valuation curve reset 15x in 14 months. Bloomberg (Thursday May 22) reported Anthropic's $30B-plus round closing 'as soon as next week' at a $900-plus-billion valuation, vaulting past OpenAI's $852B March mark, with Sequoia, Dragoneer, Altimeter, and Greenoaks each writing ~$2B. NVIDIA's Q1 transcript explicitly named Anthropic alongside the hyperscalers as a Blackwell deployment customer, confirming the equity bid matches compute-side demand.

Google I/O 2026 (May 19-22) compressed a normal three-month product cycle into one keynote — Gemini 3.5 Flash (AA Intelligence Index 55.3, Terminal-Bench 76.2%, 4x faster than incumbent frontier at ~280 tok/sec), Gemini Omni Flash (native text-and-image-and-audio-and-video into grounded video out, shipping immediately in Gemini app / Flow / YouTube Shorts), Antigravity 2.0 standalone agent IDE with managed Linux sandboxes, and TPU 8t/8i dual-chip 8th gen. Combined with OpenAI Codex 'Goal mode' GA + macOS Appshots (May 21), Anthropic's self-hosted Managed Agent sandboxes + MCP tunnels (May 19), and xAI Grok Build (W20), the strategic battleground shifted decisively from model intelligence to agent-runtime monetization and ecosystem lock-in.

The specialist canopy kept widening at a fraction of frontier cost. Cohere Command A+ shipped the first-ever Apache 2.0 frontier-adjacent MoE (218B/25B-active, runs on 2x H100, τ²-Bench Telecom jumped 37%→85% generation-over-generation). Microsoft Research Fara1.5-27B (open weights, fine-tuned on Qwen3.5) hit 72% Online-Mind2Web — beating OpenAI Operator (58.3%) and Gemini Computer Use (57.3%) — collapsing the cost structure for browser-agent fleets. Alibaba Qwen3.7-Max (May 20) became the first Chinese model in the AA Intelligence Index top 5 (56.6, ahead of Gemini 3.5 Flash) with a demonstrated 35-hour autonomous tool-use run.

The one bullet dodged: Samsung-union talks reached a tentative wage agreement late Wednesday May 20 — about an hour before the planned 18-day general strike — granting a 10.5% performance bonus pending ratification (vote May 27-28). The W20 downside scenario (Vera Rubin HBM4 supply at acute risk) is largely off the table conditional on the vote passing. So-what for the four audiences: boards should retire 'AI capex peak' theses and treat federal-procurement + capability-gating as durable vendor-risk dimensions; investors should treat the Anthropic close (not the headline valuation) as the binary read of the week and price NVIDIA on supply-not-demand; architects and operators should pull agent-runtime selection into procurement bake-offs now — model swaps inside someone else's harness are constrained.

JevonsMetcalfeGilderSoftwareJevonsHardwareHuangNetworkingMetcalfe + Gilder

The three lenses

What moved this week, and what to do about it.

13 events across the flywheel — 5 software, 4 hardware, 4 networking.

Software.

  • Google I/O 2026: Gemini 3.5 Flash GA (AA Intelligence Index 55.3, Terminal-Bench 2.1 76.2%, ~280 tok/sec — beats prior-gen Gemini 3.1 Pro on most agent benchmarks at 4x speed; list price $1.50/$9 per Mtok), Gemini Omni Flash (native multimodal create-and-edit with grounded video out, immediately in Gemini app + Flow + YouTube Shorts), Antigravity 2.0 standalone agent IDE with managed Linux sandboxes, TPU 8t/8i dual-chip 8th gen, and Managed Agents API

    blog.google/innovation-and-ai/technology/developers-tools/google-io-2026-developer-highlights, deepmind.google model card, artificialanalysis.ai, simonwillison.net

  • Alibaba Cloud Summit ships Qwen3.7-Max (AA Intelligence Index 56.6 — first Chinese model in the global top 5, ahead of Gemini 3.5 Flash; 1M context, demonstrated 35-hour autonomous 1,158-tool-call run), Qwen3.7-Plus-Preview (multimodal sibling), homegrown Zhenwu M890 AI chip (3x predecessor), and a rebuilt full-stack agentic cloud platform

    alibabacloud.com/blog/alibaba-unveils-new-ai-chip-flagship-model-and-rebuilt-cloud-stack-ai-for-agentic-era, marktechpost.com, Artificial Analysis

  • Cohere ships Command A+ under Apache 2.0 — first-ever fully-permissive license from Cohere on a 218B/25B-active MoE that runs on 2x H100s; τ²-Bench Telecom jumped 37% (Command A Reasoning) to 85%, Terminal-Bench Hard 3% to 25%, AA-Omniscience Non-Hallucination #1 at 86%; 48 languages with citation grounding for sovereign / on-prem RAG deployments

    cohere.com/blog/command-a-plus, venturebeat.com, artificialanalysis.ai, MarkTechPost

  • Microsoft Research Fara1.5 family (4B / 9B / 27B, open weights, fine-tuned on Qwen3.5) — 27B variant hits 72% Online-Mind2Web, beating OpenAI Operator (58.3%) and Gemini 2.5 Computer Use (57.3%); released with MagenticLite sandboxed browser plus FaraGen1.5 synthetic-data pipeline; collapses per-task cost for browser-agent fleets

    microsoft.com/en-us/research/articles/fara1-5-computer-use-agent, decrypt.co, Azure AI Foundry

  • Anthropic publishes Project Glasswing initial update — Claude Security beta (Opus 4.7-powered) patched 2,100 vulnerabilities in three weeks; Anthropic reiterates Mythos will NOT release publicly until 'far stronger safeguards' exist despite UK AISI showing Mythos Preview at 73% on expert CTFs and first to autonomously complete the 32-step 'Last Ones' corporate intrusion (3 of 10 runs)

    anthropic.com/research/glasswing-initial-update, aisi.gov.uk evaluation-of-claude-mythos-previews-cyber-capabilities

What this means

Two simultaneous arcs: (1) speed-intelligence Pareto frontier reset — Gemini 3.5 Flash now beats prior-gen Pro at 4x throughput, so architects should re-baseline cost-per-task this quarter rather than waiting for Pro-tier price drops; (2) agent-platform lock-in becomes the procurement decision — every flagship this week (Antigravity 2.0, Codex Goal mode, Claude Managed Agents sandboxes, Grok Build, Command A+) is engineered for sustained agentic loops, with the harness (AGENTS.md, SKILL.md, sandbox runtime) now the multi-year commit rather than the model. Boards should treat federal procurement and capability-gating (Anthropic's continued Mythos non-release; METR's first Frontier Risk Report May 19) as durable vendor-risk dimensions, not capability-fade speculation.

Hardware.

  • NVIDIA Q1 FY27: revenue $81.6B (+85% YoY), Data Center $75.2B (+92% YoY, +21% QoQ), Q2 guided to $91B (~$4B above Street, zero China DC compute assumed); supply commitments stepped to $145B; $80B buyback authorized; dividend raised 25x; Huang says Vera Rubin (NVL144, 3.6 EFLOPS FP4, Q3 2026 production) 'off to a tremendous start' and NVIDIA 'will be supply-constrained through the entire life of Vera Rubin'

    NVIDIA IR press release, SEC 8-K q1fy27pr, CFO commentary, CNBC, Reuters, Motley Fool transcript

  • Samsung Electronics and the National Samsung Electronics Union reach a tentative wage agreement at ~22:30 KST — about an hour before the planned 18-day general strike (May 21 to June 7, ~48,000 workers, ~40% global DRAM share at risk) — granting a 10.5% special semiconductor performance bonus with no cap; outcome conditional on ratification vote May 27-28; Yonhap industry estimates the strike would have removed 3-4% of global DRAM and 2-3% of NAND supply and directly threatened HBM4 ramp for Vera Rubin

    Yonhap, Korea JoongAng Daily, AJU Press, Seoul Economic Daily, Reuters via Investing.com

  • AMD announces >$10B in Taiwan ecosystem investments tied to Helios rack-scale platform (Instinct MI450X + 6th-gen EPYC 'Venice' CPUs + Pensando 'Vulcano' NICs) on track for multi-gigawatt deployments H2 2026; introduces EFB-based 2.5D packaging for Venice; confirms ODM build-out with Sanmina, Wiwynn, Wistron, Inventec — Helios positioned as the only credible second source to Vera Rubin in the same shipment window

    AMD IR press release via MarketScreener, AMD on X (May 21), Wccftech, Technosports

  • Intel 'Crescent Island' inference accelerator PCB leak: single Xe3P die paired with 20 LPDDR5X packages for 160GB at ~0.94 TB/s on a 640-bit bus, explicitly skipping HBM to sidestep the HBM4 supply crunch; H2 2026 customer sampling, volume early 2027 — first mainstream data-center AI accelerator designed around LPDDR5X as a supply hedge against the SK Hynix / Samsung HBM4 duopoly NVIDIA confirmed on Q1 call

    Wccftech, TweakTown, LavX News, original PCB leak by YuuKi_AnS / @M0_R4_H4

What this means

Architects and operators should now treat HBM4 + CoWoS allocation, not announced TFLOPS, as the binding spec for fleet density through 2027: NVIDIA's $145B supply commitments, $91B Q2 guide, and Huang's 'supply-constrained through the entire life of Vera Rubin' push the bottleneck firmly into back-end packaging, and the Samsung tentative agreement keeps that bottleneck one rejected ballot from re-tightening on May 27-28. For boards and investors, Huang's $200B Vera CPU TAM plus an $80B buyback and 25x dividend hike is the Huang flywheel re-leveraging from product to balance sheet — scarcity, not speed, is now the moat heading into AMD Helios' H2 2026 ramp and Intel Crescent Island's HBM-bypassing sampling. Custom-silicon share trajectory remains intact (TrendForce reaffirms ASIC +44.6% CAGR vs GPU +16.1%).

Networking.

  • NVIDIA Q1 FY27 segment-reporting split: Data Center Networking revenue $14.8B (+199% YoY, +35% QoQ) — InfiniBand, Spectrum-X, NVLink combined; multi-year silicon-photonics agreements with Coherent, Corning, and Lumentum reaffirmed alongside the Marvell NVLink Fusion partnership

    Data Center Knowledge, NVIDIA investor materials, NVIDIA CFO commentary

  • EBO MSA membership expands from 17 to 23 vendors with HPE (marquee server / switch signature), Bellwether, JPC Connectivity, Mixx Technologies, TIME Interconnect, and TFC Communication joining the AMD / Cisco / Meta / Oracle-led expanded-beam optical connector standard

    ebomsa.org, SDxCentral, Converge Digest, 3M Newsroom

  • Lightmatter unveils Guide DR liquid-cooled in-chassis Laser NIC in OCP NIC 3.0 form factor at InterConnect 2026 — up to 51.2 Tbps per module and 204.8 Tbps per 1RU switch tray; relocates laser sources from faceplate to chassis to break the CPO / NPO scale-up faceplate bottleneck; samples Q4 2026

    Lightmatter press release, Converge Digest, The AI Journal

  • Google I/O keynote concentrated entirely on Gemini 3.5 Flash / Omni Flash and Antigravity 2.0 — no MRC adoption confirmation, no Jupiter / Apollo update, no proprietary fabric alternative disclosed, leaving W19 prediction p13 (second-hyperscaler MRC at >50K-GPU scale) untested at Google's biggest annual stage

    Google Developers Blog, Google I/O 2026 keynote, blog.google

What this means

Architects and capex planners should now treat optical interconnect as a peer procurement category to GPUs rather than a GPU attachment: NVIDIA's $14.8B networking quarter (+199% YoY) plus HPE's signature on the EBO MSA put scale-across optics on the same line item as accelerators, while Google's fabric silence at I/O leaves the open MRC + EBO + UEC stack as the de facto hyperscaler standard by default. The Gilder networking-to-software arc is now legible in dollars — bandwidth spend is compounding faster than compute inside the AI factory, and 2026-2027 designs should budget optics as a separately-clearable, multi-vendor stream rather than a NIC-and-cable line buried under GPU.

Capital flow

Money in, revenue out.

4 categories tracked. Capital deployment up in 2 of 4; revenue follows at multiples of 0.21 to 0.6.

The four-category scorecard. Where capital is going in, where revenue is coming out, and how much of it is real. The one chart for the boardroom.

  • Frontier Labs

    OpenAI, Anthropic, Google DeepMind, xAI

    Capital In

    ~$52B ↑

    vs ~$52B

    Revenue Out

    ~$48B →

    vs ~$48B

    Burn / Rev

    0.9

    Movement

    Anthropic's $30B-plus round at $900B-plus valuation moved from 'in talks' (Bloomberg May 12) to 'closing as soon as next week' (Bloomberg May 22) — Sequoia, Dragoneer, Altimeter, Greenoaks each writing ~$2B; vaulting past OpenAI's $852B March mark. Boards holding 'AI capex peak' theses should retire them this week; investors should treat the close (not the headline valuation) as the binary read.

  • Hyperscaler-Hosted

    Azure-OpenAI, AWS-Anthropic, Google Cloud-Gemini, Oracle-OCI

    Capital In

    ~$60B ↑

    vs ~$58B

    Revenue Out

    ~$20B ↑

    vs ~$15B

    Burn / Rev

    0.33

    Movement

    NVIDIA Q1 FY27 (May 20) printed Data Center revenue $75.2B (+21% QoQ from Q4 FY26's $62.3B, +92% YoY) — investors should retire the 'hyperscaler capex isn't translating to chip revenue' bear case; it is, demonstrably, and the Q2 guide of $91B with zero China assumed says next quarter accelerates. Operators planning 2H26 should escalate HBM4 supply assumptions immediately.

  • Neoclouds

    CoreWeave, Nscale, Crusoe, Lambda, Fluidstack, IREN

    Capital In

    ~$30B →

    vs ~$30B

    Revenue Out

    ~$3.0B →

    vs ~$3.0B

    Burn / Rev

    0.10

    Movement

    Quiet week for fresh neocloud closures, but NVIDIA Q1 FY27 transcript reframes the category: NVIDIA's $145B supply commitments + in-quarter buybacks while remaining a guaranteed buyer of CoreWeave's unsold capacity (per Reuters reporting on the $6.3B arrangement) means architects must now price the 'circular flow' risk into neocloud GPU contracts — vendor concentration is no longer just a procurement question, it's a counterparty question.

  • On-Prem / Hybrid

    Enterprise GPU clusters, sovereign and national programs, Cisco / Dell / HPE

    Capital In

    ~$45B →

    vs ~$45B

    Revenue Out

    Indirect →

    vs Indirect

    Burn / Rev

    n/a

    Movement

    Cisco's W20 print (Q3 FY26 raised FY26 AI orders guide $5B to $9B) anchors W21 on-prem demand; NVIDIA Q1 FY27 networking +199% YoY confirms the same demand wave from the chip side. PJM accelerated its proposed backstop reliability auction to September (from March, May 19) for data-center load — architects should pull network / optical line items forward in 2H26 build plans and pre-commit power before pre-committing GPU SKUs.

Burn-to-Revenue is revenue divided by committed capital. Lower means more capital is going out than coming in.

See the full breakdown

Signal vs noise

What’s real, what’s noise.

6 claims this week — 4 signal, 2 noise.

Each claim is scored 1–5 on source quality and triangulation. Anything 2 or below is flagged as noise. Where consensus is wrong, we say so.

  • 5 / 5

    NVIDIA Q1 FY27 (May 20) Data Center revenue $75.2B (+21% QoQ from $62.3B; +92% YoY); Q2 FY27 guide $91B plus-or-minus 2% with zero China DC compute assumed; total supply commitments + prepaids $145B; $80B additional buyback authorized; quarterly dividend lifted 25x ($0.01 to $0.25).

    Sources: NVIDIA investor relations, SEC 8-K (q1fy27pr), Motley Fool transcript, MarketScreener CFO commentary

    Real and decisive. Strongest possible primary-source signal that hyperscaler + AI-cloud + sovereign capex is still translating into chip revenue, and supply not demand is the 2H26 binding constraint. Would change the read only if Q2 prints below ~$86B Street consensus or if a clear China DC compute upside emerges and isn't ramped (signaling demand-side pressure).

  • 4 / 5

    Anthropic to close $30B-plus round at $900B-plus valuation as soon as the week of May 25, vaulting past OpenAI's $852B March mark; Sequoia / Dragoneer / Altimeter / Greenoaks each ~$2B; Founders Fund + General Catalyst participating.

    Sources: Bloomberg (Mascarenhas / Torrence / Ghaffary May 22), Moneycontrol / EconoTimes syndication

    Real but not yet closed in W21. Sourcing is Bloomberg primary plus syndication; round reported by people with direct knowledge but commitments 'still being finalized.' Would change the read if the close slips into mid-June (signal that bond-market term-sheet pressure is biting) or if Microsoft / Salesforce join as strategic allocators (would change the antitrust risk profile).

  • 4 / 5

    Samsung Electronics and union reached a tentative wage agreement May 20 ~22:30 KST — about an hour before the planned 18-day strike (May 21 to June 7, 48K workers, ~40% global DRAM share at risk); 10.5% special semiconductor performance bonus with no cap; outcome conditional on ratification vote May 27-28; strike would have removed 3-4% global DRAM and 2-3% NAND supply and directly threatened HBM4 ramp for Vera Rubin.

    Sources: Yonhap, Korea JoongAng Daily, AJU Press, Seoul Economic Daily, Reuters via Investing.com

    Real and conditional. Tentative agreement is primary-sourced; ratification vote is binary. Operators planning H2 capacity should hold HBM allocation triage until the vote outcome is reported May 28-29. A ratification rejection would reactivate the W20 supply-risk thesis on three days' notice.

  • 4 / 5

    Cohere ships Command A+ under Apache 2.0 — first-ever fully-permissive license from Cohere on a 218B/25B-active MoE that runs on 2x H100s; τ²-Bench Telecom 37% to 85% generation-over-generation; AA-Omniscience Non-Hallucination #1 at 86%; 48 languages with citation grounding.

    Sources: cohere.com/blog/command-a-plus, venturebeat.com, artificialanalysis.ai, MarkTechPost

    Real and structural. First Apache 2.0 frontier-adjacent MoE from a Western lab — sovereign-AI and data-residency procurement teams now have an on-prem option that previously required a closed API or a license-restricted open-weights model. Would change the read only if Cohere reverses the license posture on a subsequent model release, or if a major hyperscaler ships an equally permissive frontier MoE first.

  • 2 / 5 — noise

    Custom-silicon ASIC share of AI server shipments will reach ~40% by 2030, with hyperscaler in-house silicon (TPU, Trainium, MTIA, Maia) collapsing NVIDIA inference market share to 20-30% by 2028.

    Sources: TrendForce March 2026, Tom's Hardware May 2026, Bloomberg Intelligence (via Introl); New Street model output reproduced in derivative coverage

    Plausible directionally; the +44.6% YoY ASIC growth vs +16.1% GPU is well-sourced. But the 20-30% NVIDIA inference share by 2028 is a New Street model output — vendor-friendly framing for ASIC co-designers. Thesis requires hyperscalers actually shipping at scale; historically those programs slip. Would change the read if AWS, Google, or Meta publish per-rack TCO numbers proving 65% TCO inference advantage in production.

  • 1 / 5 — noise

    Microsoft Build 2026 (May 19-22) delivered new MAI / capex / RPO color this week.

    Sources: Inherited from W20 watchlist framing; not corroborated by any in-window primary source

    Hype / misattribution. Microsoft Build 2026 is scheduled June 2-3, 2026 (per Microsoft Foundry blog session schedule), NOT May 19-22 — that window is Google I/O, not Build. The W21 Microsoft news was Surface for Business hardware launch (May 19) and Microsoft India data-center commentary, not capex / RPO disclosures. Watch W23 (June 2-3) for the actual Build content.

Early warning panel

The levers we monitor.

10 metrics tracked — 3 rising, 0 falling, 7 steady.

Current vs prior period. Each metric has a threshold where the read materially changes — this panel flags the inflection before it lands in headlines. Click any metric for the methodology and this-week read.

  • Frontier lab cash position (avg months runway, top 3)

    ~30 mo↑vs ~30 mo

    Threshold: <18 mo triggers re-rating risk

    What this measures

    Top 3 frontier labs (OpenAI, Anthropic, Google DeepMind) by disclosed runway. Anthropic's $30B-plus / $900B-plus round (Bloomberg May 22, target close end-May) will materially extend top-3 average once funded. OpenAI's April 27 capped revshare amendment ($38B cap, non-exclusive license through 2032) further reduces forward burn pressure. Carry W20 figure flat pending Anthropic close; boards should not assume frontier-lab funding pressure as a forcing function for short-term commercial concessions.

  • Hyperscaler capex / AI revenue ratio (top 4 weighted)

    ~5.0-5.2→vs ~5.0-5.2

    Threshold: >6.0 invites investor pushback at next earnings

    What this measures

    Top 4 hyperscalers (MSFT, GOOG, META, AMZN) weighted aggregate of total capex divided by AI-attributable revenue. Microsoft FY26 capex tracking ~$190B against ~$37B AI ARR ≈ 5.1x; Google capex ~$180-190B per Pichai I/O keynote. NVIDIA Q1 FY27 hyperscale revenue $37.9B confirms the spend is real chip revenue. Investors monitoring 'capex bubble' should switch the hypothesis to power / HBM4 supply constraints, not demand.

  • CoreWeave revenue backlog

    $99.4B→vs $99.4B

    Threshold: Conversion velocity matters more than gross figure

    What this measures

    Booked but unrecognized revenue. May 7 audited Q1 figure unchanged in W21; next print is Q2 in early August. Operators evaluating neocloud counterparty risk should re-read NVIDIA's $6.3B 'guaranteed buyer of unsold capacity' arrangement as the relevant downside-floor structure for the category. Conversion velocity remains the metric to actually watch.

  • NVIDIA Q-over-Q data center revenue

    $75.2B (Q1 FY27)↑vs $62.3B (Q4 FY26)

    Threshold: +21% QoQ; Q2 FY27 guide $91B implies further +21% QoQ

    What this measures

    W21's defining datapoint. Q1 FY27 print on May 20: Data Center segment $75.2B (+21% QoQ from Q4 FY26's $62.3B, +92% YoY); Q2 guide $91B plus-or-minus 2% with zero China DC compute assumed. Architects modeling 2H26 should re-baseline NVIDIA DC revenue trajectory upward; HBM4 supply (Samsung walkout averted May 20 pending ratification May 27-28) is the binding constraint, not demand.

  • Open vs closed gap on SWE-Bench Pro (coding)

    Closed +19pp→vs Closed +19pp

    Threshold: Sustained open lead reshapes enterprise procurement

    What this measures

    Top-of-board (Claude Mythos Preview 77.8%) vs top open (Kimi K2.6 / GLM-5.1 ~58.4-58.6%) ≈ 19.2pp unchanged. But second-tier compression is now notable: GPT-5.5 closed 58.6% sits at Kimi K2.6 open 58.6%. Architects evaluating self-host options should pilot Kimi K2.6 / GLM-5.1 / Cohere Command A+ against GPT-5.5 before signing multi-year closed-model contracts; the gap is ~zero outside the Mythos top spot.

  • Sovereign AI commitments (count / aggregate $)

    12 / ~$80B+→vs 12 / ~$80B+

    What this measures

    No fresh sovereign deal closed inside W21. Stargate UAE first chip shipment confirmed May 8 (UAE ambassador statement); 200 MW Q3 2026 launch on track per Emirates Insight. Riyadh Compute Pact one-year retrospective May 5 cited 23 cross-border data-sharing agreements activated in Year 1, 14 with a UAE leg. Operators with EMEA workloads should now treat GCC sovereign compute as a credible alternative landing zone.

  • PJM 2026/27 capacity auction price ($/MW-day)

    $329.17→vs $329.17

    Threshold: 11x in 24 months — power is the new binding constraint

    What this measures

    The 2026/27 BRA cleared July 2025 at $329.17/MW-day, unchanged. PJM May 19 backstop reliability auction acceleration (March to September) signals data-center load growth is forcing structural responses; capacity prices in PJM stay cap-bound through at least 2028/29 BRA. Architects should not assume any near-term price relief from forward auctions; budget capacity at-cap through 2028.

  • Time-to-power, busiest US markets (months)

    60-84 (new PJM); 36-48 (existing PJM queue)→vs 60-84 (new PJM); 36-48 (existing PJM queue)

    What this measures

    Months from new-load interconnection request to energization. PJM April 29 announced reopened queue (811 projects, 220 GW) with target review under 2 years using Tapestry / Google AI tools — directionally supportive but total time-to-power unchanged. Constellation Energy May earnings flagged data-center customers pausing decisions to see how PJM colocation / backstop rules shake out. Architects should pre-commit power before pre-committing GPU SKUs.

  • Cost-per-task, frontier reasoning model

    ~$0.10-$0.15 (effective, with hidden reasoning tokens)→vs ~$0.10-$0.15 (effective, with hidden reasoning tokens)

    Gemini 3.5 Flash GA at I/O priced 3x prior Flash list but 'half per-task'

    What this measures

    Median cost across frontier-tier reasoning models for a benchmark complex task. Gemini 3.5 Flash GA (May 19) priced $1.50/$9 per Mtok — 3x prior Flash list price, but Pichai claim of 'frontier capabilities at less than half the price' refers to per-task vs Mythos / GPT-5.5, not vs prior Flash. Operators running agents at scale should re-benchmark on cost-per-task (e.g., RCB) not list price; cache-aware prompting still cuts realized cost ~22% median.

  • Custom silicon share of incremental AI compute

    ~33-36%↑vs ~33-36%

    Threshold: >35% materially compresses merchant GPU pricing

    What this measures

    TrendForce 2026 outlook reaffirmed: ASIC AI servers 27.8% of shipments (multi-year peak), GPU 69.7%; ASIC growth +44.6% YoY vs GPU +16.1%. NVIDIA Q1 FY27 muted stock reaction (-1.6% AH despite blowout) explicitly priced the custom-silicon erosion arc. Alibaba's Qwen3.7-Max ships on homegrown Zhenwu M890 (3x predecessor); Google announced TPU 8t/8i 8th-gen. Investors with >20% NVIDIA exposure should diversify into ASIC co-design (Broadcom, Marvell — together ~95% of co-design market) and advanced packaging / power tier (TSMC, Vertiv).

Predictions

What we expect next.

5 predictions for the next 30-90 days, confidence 60%-80%.

Each prediction is falsifiable, time-bounded, and tied to a specific signal we will watch. Future issues score these hit, miss, partial, or pending and build a public track record.

Prediction 01

70%

confidence

Hardware

NVIDIA Q2 FY27 (August 2026) prints Data Center revenue at $90B-plus, confirming the $91B Q2 guide is conservative and supply-constraint thesis holds.

Deadline: By August 31, 2026

Trigger: NVIDIA Q2 FY27 earnings release; HBM4 supply commentary post-Samsung ratification vote; hyperscaler Q2 capex prints.

Prediction 02

80%

confidence

Capital

Anthropic round closes at a final post-money valuation between $900B and $1.1T with at least three of (Sequoia, Dragoneer, Altimeter, Greenoaks) named as lead investors publicly.

Deadline: By June 13, 2026

Trigger: Anthropic primary disclosure; SEC Form D if applicable; lead-investor naming via Bloomberg / Reuters / The Information.

Prediction 03

60%

confidence

Software

At least one Fortune 500 enterprise discloses standardization on a single agent-platform runtime (Claude Code / Codex / Antigravity 2.0 / Grok Build) for greater than $50M annualized commitment.

Deadline: By August 31, 2026

Trigger: Enterprise architecture announcements; vendor earnings color from OpenAI partnership disclosures, Anthropic, Google Cloud, xAI; F500 case studies.

Prediction 04

70%

confidence

Networking

At least three of (Marvell, Broadcom, NVIDIA networking, Arista) post Q2 networking-segment revenue growth greater than 40% YoY, validating EBO MSA momentum into Q2 prints.

Deadline: By August 31, 2026

Trigger: Q2 earnings prints from networking-focused vendors; NVIDIA Q1 already at +199% YoY sets the bar.

Prediction 05

60%

confidence

Software

Gemini 3.5 Pro launches by July 15, 2026, scoring AA Intelligence Index greater than or equal to 60, taking the closed-frontier intelligence lead from GPT-5.5 (currently 60.2).

Deadline: By July 15, 2026

Trigger: Google blog / DeepMind announcement; Artificial Analysis leaderboard update; vendor benchmark releases.

Track record

Scoring prior predictions.

5 prior predictions: 0 hit, 0 miss, 2 partial, 3 pending. Hit rate 0%.

5 predictions across issues so far. Hit rate: 0%. Hits 0, misses 0, partials 2, pending 3.

Prediction 01

55%

confidence

Hardware

Samsung HBM4 shipments to NVIDIA Vera Rubin drop by more than 25% in May vs April due to the May 21-June 7 walkout, with at least one downstream hyperscaler publicly delaying a Q3 capacity ramp.

Deadline: By July 31, 2026

Trigger: TrendForce / SemiAnalysis May-June supply data; NVIDIA Q1 FY27 commentary May 20; hyperscaler Q2 capex / supply commentary.

partial— Samsung-union tentative agreement May 20 averted the 18-day walkout; ratification vote May 27-28. HBM4 shipments likely NOT to drop >25%; this prediction is on track to MISS. Confidence revised down. Score moves to miss if the ratification passes; score moves to hit only if the ratification fails and the strike actually proceeds.

Prediction 02

65%

confidence

Software

Anthropic discloses or third-party developer surveys document a greater than 20% drop in active Claude Code third-party agent users within 30 days of the June 15 pricing cutover.

Deadline: By August 15, 2026

Trigger: Anthropic earnings / blog commentary; GitHub stars / npm downloads on Claude Agent SDK; third-party agent vendors (OpenClaw, T3 Code, Conductor, Zed) public statements.

pending— Anthropic Glasswing update (May 22) reiterated Mythos non-release stance; Agent SDK June 15 cutover not yet active. Anthropic Managed Agents sandboxes + MCP tunnels (May 19) may absorb some third-party agent workload. Pending.

Prediction 03

60%

confidence

Software

At least one Fortune 500 enterprise discloses a production deployment greater than $10M annualized of a specialist video / embodied multimodal model (Perceptron Mk1, MolmoAct 2 successors, or analogous) as a primary tier rather than a general LMM.

Deadline: By September 30, 2026

Trigger: Vendor case studies; enterprise architecture announcements; quarterly earnings color from Perceptron, Ai2 partners.

pending— Cohere Command A+ (Apache 2.0 frontier-adjacent) and Microsoft Fara1.5 (open-weight browser-agent specialist) both shipped this week, expanding the addressable open-weights specialist market. No F500 procurement disclosure yet — pending.

Prediction 04

70%

confidence

Capital

Cisco's Q4 FY26 AI orders confirm the $9B annual run rate (Q4 AI orders at least $2.5B), validating the enterprise AI fabric demand-pull narrative.

Deadline: By August 20, 2026

Trigger: Cisco Q4 FY26 earnings release (early August).

pending— Cisco Q4 FY26 print expected early August. NVIDIA Q1 FY27 networking +199% YoY confirms the demand wave from the chip side — Cisco's $9B FY26 AI orders guide raise looks well-supported. On track.

Prediction 05

60%

confidence

Networking

Expanded Beam Optical MSA publishes a v1.0 spec within 90 days of launch (May 12), with at least one in-production deployment announced by a hyperscaler member (AMD, Cisco, Meta, Oracle).

Deadline: By August 12, 2026

Trigger: ebomsa.org publications; member vendor blogs; OCP Future Technologies Symposium filings.

partial— EBO MSA membership expanded 17 to 23 vendors May 18 (HPE marquee addition, Bellwether, JPC Connectivity, Mixx, TIME, TFC). v1.0 spec not yet published. Member growth is positive signal but spec + in-production deployment still pending. On track.

Watchlist

On the radar this week.

6 catalysts to watch, starting May 25 - 30.

Specific catalysts that would change the read materially. Watching these tells us whether the thesis is strengthening or weakening.

  • May 25 - 30

    Anthropic round close — final terms, lead investor, full investor list

    Bloomberg May 22 reported close 'as soon as next week' at $30B-plus / $900B-plus. Tests prediction p23; reset frontier-lab valuation curve. A slip past end-May or a haircut below $700B would be the loudest re-rating signal in a year.

  • May 27 - 28

    Samsung-union ratification vote on tentative HBM4 agreement

    Binary outcome. Approval keeps the 18-day walkout off the table and HBM4 supply for Vera Rubin intact; rejection re-activates the W20 supply-risk thesis on three days' notice and tests prediction p17 the other way.

  • June 2 - 3

    Microsoft Build 2026 (San Francisco) — MAI-line announcements + Foundry updates

    Microsoft notably quiet through I/O week. Build is where MAI-line frontier models, Copilot agent stack moves, and Foundry consolidation are likely to land. Procurement teams running Azure / Foundry should hold model-selection decisions until after Build.

  • June 1 - 5

    Computex Taipei 2026 + NVIDIA GTC Taipei (Huang keynote June 1)

    Computex + NVIDIA GTC Taipei deliver the H2 2026 silicon roadmap (Vera Rubin partner ecosystem, Helios shipping milestones, Crescent Island OEM commitments). First Computex with Saturday-cadence reporting — fully inside W22 window.

  • June 9 - 13

    Apple WWDC 2026 — on-device model strategy, Apple Intelligence v2

    Apple absent from the model-launch cycle through Q2. WWDC is the only major Apple model-disclosure window in 1H. Mobile / edge architects building on-device AI should expect significant Apple Intelligence repositioning and possibly partnered closed-model integration changes.

  • May 26 - July 15

    Gemini 3.5 Pro launch — frontier intelligence rematch

    Google promised Pro 'next month' (June 2026). This is the model that determines whether Google retakes the AA Intelligence Index lead from GPT-5.5 (currently 60.2) or stays on the second podium tier. Tests prediction p26. Buyers planning Pro-tier procurement should hold pending Pro pricing and benchmarks.

Companion reads

The rest of the spine.

The AI Stack Weekly is the cross-stack flywheel read. Pair it with the model-and-tree spine and the working framework to get the full picture.

Edits this issue

  • — First fully Saturday-cadence publish via the auto-scaffold cron (the cron itself fired late on its first try due to GHA queue lag + a permission gap; both fixed for next week). Window: Mon May 18 - Sat May 23, 2026.
  • — NVIDIA Q-over-Q data center revenue lever updated with the Q1 FY27 print: $62.3B (Q4) to $75.2B (Q1 FY27), +21% QoQ, +92% YoY. The lever's threshold reframed against the Q2 $91B guide.
  • — W20 priors scored: p17 (Samsung walkout) moves to partial trending miss after the May 20 tentative agreement; p21 (EBO MSA spec) moves to partial after the membership expanded 17 to 23 vendors.
  • — Editorial correction: W20 watchlist mis-dated Microsoft Build as May 19-22; Build is actually June 2-3. Surfaced as a Score-1 noise call in signalVsNoise and added to W21 watchlist correctly.

About this brief

Compiled from public announcements, SEC filings, earnings transcripts, and official lab and vendor publications. Every quantitative claim is graded 1–5 on source quality. Claims graded 2 or below are flagged as noise. The thesis the brief defends is published separately and updated only when a hypothesis materially changes.

Authorship

Written by Brian Letort. Independent analysis. All sources cited are public. Not investment guidance.

Operate. Publish. Teach.