Edition 11 · May 25–31, 2026
Last week was distribution — Search agents, managed sandboxes, Glasswing tallies. This week is unit economics at scale. On May 28, Anthropic closed $65 billion in Series H funding at a $965 billion post-money valuation, reporting run-rate revenue above $47 billion and signing multi-gigawatt compute agreements with Amazon, Google/Broadcom, and SpaceX in the same breath as Claude Opus 4.8. Meanwhile Jensen Huang told Taipei that NVIDIA now spends roughly $100 billion annually in Taiwan — heading toward $150 billion — while breaking ground on the Constellation campus. COMPUTEX opens with Vera Rubin and an Arm laptop SoC (N1X) on deck; Intel ships purpose-built Arc G3 handheld silicon; SoftBank pledges €75 billion to French AI infrastructure. Capital, watts, and wafers — not demos.
May 28: Anthropic’s primary release states the round was led by Altimeter, Dragoneer, Greenoaks, and Sequoia, with co-leads including Capital Group, Coatue, D1, GIC, ICONIQ, and XN. The post notes $15 billion of the total comprises previously committed hyperscaler investments — including $5 billion from Amazon announced in April — and names strategic partners Micron, Samsung, and SK hynix alongside the financial syndicate.
$65B
Series H proceeds (primary)
$15B
Hyperscaler commitments (incl. Amazon)
$380B → $965B
Valuation step since Series G (Feb 2026)
Anthropic also states Claude is available on AWS, Google Cloud, and Microsoft Azure — the first frontier model family on all three hyperscalers — with AWS remaining the primary training partner.
Anthropic — Series H TechCrunch — Series HThe Series H post is explicit about infrastructure, not just valuation. Anthropic lists: up to 5 GW of new capacity with Amazon; 5 GW of next-generation TPU capacity with Google and Broadcom; and GPU access through SpaceX in Colossus 1 and Colossus 2. For procurement teams, the story is binding future watts against enterprise Claude demand — not benchmark slides.
Shipped · May 28
claude-opus-4-8
$5 / $25 per M tokens
unchanged vs Opus 4.7
Anthropic’s Opus 4.8 post positions the release as a modest but tangible upgrade — notably candid in a hype-heavy week. Shipped capabilities include Dynamic Workflows (research preview in Claude Code): plan work, run hundreds of parallel subagents, verify outputs before merge. Effort control on claude.ai lets users trade latency for depth. Fast mode runs at ~2.5× speed with pricing at $10/$50 per million tokens — Anthropic says fast mode is now three times cheaper than on prior Opus generations.
“Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.”
— Anthropic Opus 4.8 announcement, citing internal evaluations
Last edition quantified Project Glasswing vulnerability volumes. This week’s Opus 4.8 post adds the forward chapter: Anthropic says it is making swift progress on cyber safeguards and expects to bring Mythos-class models to all customers in the coming weeks — while noting no lab yet has safeguards strong enough for unconstrained release. Treat “weeks” as guidance, not a ship date; enterprise Claude Security remains the interim defensive surface.
Glasswing initial update — 10k+ partner-reported critical findings
Opus 4.8 — Mythos public release “coming weeks” if safeguards hold
General Mythos-class availability — monitor system cards + access tiers
May 27: Speaking in Taipei, Jensen Huang said NVIDIA spends about $100 billion per year in Taiwan today and is on a path toward $150 billion — up from roughly $10–15 billion four to five years ago per wire and trade-press accounts. The same visit marked a launch celebration for NVIDIA Constellation, the company’s first overseas headquarters (Beitou-Shilin Technology Park), targeting operations around 2030 with on the order of 4,000 direct employees in company statements relayed by press.
Nikkei Asia — Huang in TaipeiAnalysts note the $150B figure flows through existing Taiwanese partners (TSMC packaging, ODM system builds) — signaling commitment and supply-chain strain rather than a single new fab check.
Lisa Su, in Taipei ahead of COMPUTEX, told reporters AMD will invest more than $10 billion in Taiwan’s AI ecosystem — context for Venice on 2nm and MI-series GPU competition.
The Indian Express — COMPUTEX previewCOMPUTEX runs June 2–5 in Taipei, but the pre-show drumbeat dominated May 25–31 coverage. Huang’s GTC Taipei keynote is scheduled June 1, 11:00 a.m. local time (11:00 p.m. ET, May 31). Trade press expects Vera Rubin platform detail and the N1X — NVIDIA’s first laptop SoC — alongside coordinated posts from NVIDIA, Microsoft, and Arm teasing “a new era of PC.” Dell and other OEMs have embargoed handheld/laptop unveilings timed to the keynote window.
May 21 (shipping into COMPUTEX narrative): AMD’s official release states Venice — 6th Gen EPYC on Zen 6 — is ramping in Taiwan on TSMC’s 2nm node, with future Arizona ramp planned. AMD calls it the first HPC product in the industry to enter production on that node. Press and analyst writeups cite up to 256 cores and materially higher memory bandwidth versus Turin — check AMD’s tables for final TDP and clock disclosures.
AMD IR — Venice rampIntel’s newsroom introduced Arc G3 and Arc G3 Extreme — 14-core Panther Lake derivatives with Arc B370/B390 graphics — built for Windows 11 handhelds rather than repurposed laptop silicon. Named partners include Acer Predator Atlas 8, MSI Claw 8 EX AI+, and OneXPlayer, with OEM systems rolling from June 2026 and broader availability through the year. NPUs on-die support Copilot+ when docked — edge AI meets portable gaming form factors.
Intel Newsroom — Arc G-Series Ars Technica — Arc G3May 31: SoftBank announced a record €75 billion investment plan for AI-related infrastructure in France, with €45 billion targeted by 2031 for data centers in Hauts-de-France (Loon-Plage, Bouchain, Bosquel). France’s 95% decarbonized electricity mix is central to the pitch; Schneider Electric partners on design and equipment for an initial 3 GW capacity scalable toward 5 GW.
France in English — SoftBank planMay 28 reporting (The Information, relayed by Reuters and trade press) says Microsoft will showcase homegrown AI models at Build (June 2–3), led by a coding model meant to reduce GitHub Copilot’s reliance on OpenAI and Anthropic APIs. Treat model names and benchmarks as undisclosed until keynote — the strategic signal is multi-model Copilot orchestration, not a single partner lock.
Gadgets Now — Build preview Microsoft Foundry — May 2026Anthropic’s $47B run-rate and first operating-profit trajectory (cited in financial press) suggest enterprise Claude Code adoption is real revenue, not just GPU circularity. Multi-gigawatt contracts pre-buy the inference curve.
$965B private marks and $150B Taiwan spend narratives can coexist with memory bottlenecks and geopolitical export friction. Until Mythos ships with safeguards, offensive capability remains concentrated — Glasswing bought time, not equilibrium.
Anchors immediately after this edition — confirm times in primaries before locking travel or compliance calendars.
Livestream expected — Vera Rubin, N1X, and partner OEM reveals; hands-on demos at COMPUTEX booths from June 2.
Agent Framework GA, Foundry multi-agent sessions, rumored first-party coding model — catalog at build.microsoft.com.
Keynotes from AMD, Intel, Qualcomm; supply-chain meetings on HBM and CoWoS capacity.
Common shareholders of record receive the $0.25 quarterly dividend (payable June 26) per FY27 Q1 materials.
Keynote June 8, 10 a.m. PT — on-device model and developer tool expectations; online + limited Apple Park.
High-risk obligations still legally slated for Aug 2 unless the May 7 Digital Omnibus is enacted — Annex III delay to Dec 2027 proposed but not yet law.