Week 22Published

Anthropic 5/28 triple-drop crushes the week: Series H $965B + Opus 4.8 + Dynamic Workflows / Jensen bets $150B on Taiwan / memory duo crosses $1T

anthropicopus-4-8dynamic-workflowsopenainvidiadeepseekkimisk-hynixfujitsuperplexityhardware
On this page

Date range: 2026-05-25 (Mon) → 2026-05-31 (Sun) Compiled: 2026-06-03 (full week now covers 5/30 + 5/31 backfill) Sources: Anthropic Newsroom, Smol AI News, NVIDIA Newsroom, CNBC, Import AI 458, PR Newswire, QbitAI, Aliyun Model Studio, and more.

🌐 Lead story (full week 5/25-31)

"Anthropic 5/28 triple-drop crushes OpenAI head-on; the structural hardware bottleneck keeps compounding."

The Mon-Wed narrative of "hardware takes the baton from models" was steamrolled on Thursday (5/28) by Anthropic's three-piece combo —

  • Anthropic Series H raises $65B at a $965B post-money valuation, run-rate revenue $47B+, the first AI company to clearly leap past OpenAI's $850B in private-market valuation.1
  • Claude Opus 4.8 released: synchronous upgrades across coding / agentic / honesty axes, same price, no hike.2
  • Dynamic Workflows: a single session can orchestrate hundreds of parallel sub-agents to complete repo-scale migrations — pushing Managed Agents into "fleet-scale" territory.
  • Same day, Claude Code v2.1.154 defaults to Opus 4.8, adds /effort xhigh, ships Fast Mode (2x price for 2.5x speed), and a new /workflows command.
  • Hardware kept compounding this week: Jensen in Taipei pledges $150B/year Taiwan purchasing + Constellation campus, SK Hynix / Micron both cross $1T the same day, Fujitsu signs OpenAI + Anthropic the same day.
  • Elsewhere: CNN sues Perplexity for copyright infringement — another heavyweight name joins the AI-vs-media docket.
  • The Chinese camp stayed quiet: DeepSeek is in final sprint for V4.1; Kimi officially retired the K2 line on 5/25 and moved everything to K2.6.

⭐ Personal Top 5 (usable today / new this week)

Ranking by: real capability lift × low onboarding cost × direct cost / productivity gain

Rank Item One-liner
🥇 1 Claude Opus 4.8 + Claude Code v2.1.154 Fast Mode Same-price Opus upgrade + /effort xhigh deep-reasoning tier + 2x price for 2.5x speed Fast Mode. Daily Claude Code users get the gain just by upgrading.2
🥈 2 Codex CLI v0.135.0 — codex doctor + Vim text-object codex doctor one-shot diagnostics for environment / auth / MCP / network — slashes debug time; Vim users finally get text-object editing.
🥉 3 /workflows command + Dynamic Workflows preview Even on personal projects you can kick off "refactor across 200 files" or "batch API migration" in Claude Code with a single sentence. The endgame form of harness engineering opens up to individuals.
4 DeepSeek V4-Pro permanent 1/4 price lands 5/31 Announced W21, takes effect this weekend, permanent — no rollback. After 5/31, all individual-developer API calls bill at the new price.
5 Kimi K2.6 single-version endpoint (K2 line retired 5/25) Individual Kimi API users must move to K2.6 — migration to be completed this week; upside is a stabler endpoint and better performance.

🏢 Team / Enterprise Top 5 (teams / enterprises can act this week)

Ranking by: security / compliance improvement × deployment scope × strategic signal strength

Rank Item One-liner
🥇 1 Anthropic Dynamic Workflows + Series H $965B valuation Dynamic Workflows makes repo-scale AI migration / refactor a workflow you can actually cut a PO against; same week, Series H passes OpenAI on valuation, cementing Anthropic as the leading "first-vendor" candidate for enterprise AI. Procurement priorities need to be re-stacked.1
🥈 2 Jensen's $150B/year Taiwan purchasing + Constellation campus A clear signal to every GPU-dependent CTO / CIO that NVIDIA's supply chain stays anchored in Taiwan for the next 3 years. "Build vs rent" decisions need to be re-modeled.
🥉 3 SK Hynix + Micron both cross $1T / HBM 2026 capacity sold out HBM is the third structural bottleneck after GPUs and the grid. New AI project hardware-procurement cycles need to be re-estimated under an "HBM scarcity" assumption.
4 Fujitsu double-signs OpenAI + Anthropic the same day The biggest sample yet for "no enterprise picks just one"; any company currently betting on a single-vendor strategy needs to re-audit its risk model.
5 CNN × Perplexity copyright suit A legal-risk inflection point for enterprises procuring or embedding Answer-Engine-class AI products. Citation policy / scraping-source compliance need an audit.10

🔵 Top Two

Anthropic / Claude Code

Date Event One-liner
05-25 Pope Leo XIV's first encyclical Magnifica Humanitas co-presented with Anthropic co-founder Chris Olah 135 years after Rerum Novarum, the Vatican issues its first encyclical aimed at the AI era. Newsletter editors read it as a symbolic event for "AI elevated to infrastructure-level concern."13
05-26 Claude Code security guidance plugin Across multiple customer workflows, security-related PR comments dropped 30-40% in practice. The most direct quantitative evidence yet for the harness engineering thesis.6
05-26 Claude Code v2.1.152 (incremental release) The v2.1.x line keeps iterating (background session handling, MCP/SDK startup, /resume, /code-review, /usage category refinements).4
05-27 Fujitsu strategic partnership Signed Japan's largest SI firm on the same day as OpenAI. See the OpenAI entry below for details.9
🔥 05-28 Series H raises $65B at $965B post-money Altimeter / Dragoneer / Greenoaks / Sequoia led; run-rate revenue disclosed at $47B+; private-market valuation passes OpenAI's $850B for the first time.1
🔥 05-28 Claude Opus 4.8 released On top of 4.7, synchronous upgrades on coding, agentic, and honesty axes; same price, no hike. /effort xhigh adds a deeper reasoning tier.2
🔥 05-28 Dynamic Workflows announced A single session can orchestrate hundreds of parallel sub-agents — repo-scale migrations / large-scale code changes go from "PoC" to "shippable workflow." Pushes Managed Agents to fleet scale.
05-28 Claude Code v2.1.154 Defaults to Opus 4.8 + /effort xhigh; adds Fast Mode (2x price for 2.5x speed); new /workflows command wires into Dynamic Workflows; slimmer system prompt.4
05-29 Claude Code v2.1.156 (hotfix) Fixes an API error on Opus 4.8 when thinking blocks were mutated.4
05-31 Claude Code v2.1.159 (infra release) Shipped late Sunday — pure internal infra cleanup, no user-visible changes. Belongs to the stable stream after the 5/28 major release.4

OpenAI / Codex

Date Event One-liner
05-25 OpenAI × Grupo Folha + Grupo UOL (Brazil) partnership 900M weekly ChatGPT users get attributed Folha de S.Paulo / UOL summaries. The news-publisher framework extends from the US / UK / France / Germany into Latin America.11
05-26 Codex CLI v0.134.0 Local session history search (case-insensitive + preview); --profile becomes the main profile selector; MCP adds OAuth + concurrent execution. The first Codex CLI that can actually search its own history.5
05-27 Fujitsu strategic partnership Signed the same day as Anthropic. The classic sample of "enterprises can't pick just one anymore."9
05-28 Codex CLI v0.135.0 Adds codex doctor (one-shot diagnostics for environment / auth / MCP / network) + /status remote connection info; Vim text-object editing lands in the TUI; Python SDK gains Sandbox presets; macOS / Zellij TUI output fixes.5
05-29 Codex CLI v0.136.0-alpha.1 Alpha preview with 139 cross-platform / cross-architecture build artifacts; no release notes — wait for the stable cut.

🟢 Chinese camp

DeepSeek

  • Ongoing | Silent sprint for V4.1: no first-party announcements this week. GitHub deepseek-ai/DeepEP (their in-house expert-parallel comms library) had a commit 5/26; awesome-deepseek-agent updated 5/23. Combined with W21's "$7B raise + V4-Pro permanent price cut," the engineering line is paving the last mile for a June V4.1 (multimodal + MCP). (repos)
  • 05-28 | Media watch: DeepSeek V4 advances together with the domestic chip stack — QbitAI industry piece (not an official release) describing V4's adaptation and flywheel acceleration on the domestic chip stack. Shows the Chinese camp pushing "cheap + sovereign compute + open source" as an integrated bundle.15
  • Media watch: Securities Times and other Chinese financial outlets continue to recap DeepSeek's "permanent price cut + mega raise" combo, reading it as the set-piece move framing H2's AI industry shakeout. (source)

Moonshot / Kimi

  • 05-25 | kimi-k2 line officially retired — Aliyun Model Studio and other channels executed per Moonshot's earlier notice; everything migrates to kimi-k2.6. A deterministic version-collapse move — unifying the K2/K2.5/K2.6 service surface to K2.6, clearing the slot for K3 (Q3/Q4) as set out in Yang Zhilin's internal letter. K3 did not ship in W21 or W22; any "K3 ships in May" claims are prediction-market or third-party leak speculation.14

Zhipu / GLM

No major first-party product or pricing updates this week.

🟠 Hardware & compute & power (the main battlefield this week)

NVIDIA / Jensen Huang

  • 05-27 | Jensen in Taipei: $150B/year Taiwan purchasing + Constellation campus — at the Taipei employee town hall he announced NVIDIA's annual Taiwan supply-chain spend jumps from $100B to $150B; called Taiwan "the epicenter of the AI revolution"; the new 4,000-seat "Constellation" HQ campus in north Taipei breaks ground end of year and opens by 2030; a roster of 150 Taiwan partners published. Under US "reshore" pressure, NVIDIA is locking in Taiwan dependency in the opposite direction.7, Taiwan News)

xAI / Tesla / Musk

  • No major news this week. The aftershocks of W21's SpaceX S-1 ($2.8B gas turbines, $1.25B/month to Anthropic, xAI burning $6.4B in 2025) continue, but no fresh first-party announcements.

Other hardware / compute / power

  • 05-27 | SK Hynix and Micron both cross $1T market cap the same day — SK Hynix +11% intraday in Seoul, market cap ~$1.1T (KRW 1,624T); Micron crossed into the trillion-dollar club the same window; global memory-chip makers combined > $3T. SK Hynix has fully sold out DRAM / NAND / HBM 2026 capacity. HBM as AI's third structural bottleneck is now officially priced in.8, Invezz: SK Hynix from $100B to $1T in 16 months)
  • 05-26/27 | In-house AI ASICs rise to 27.8% of high-end AI server share (+44.6% YoY shipments); AMD starts 256-core EPYC Venice mass production — TrendForce data: Google TPU, AWS Trainium and other in-house silicon are clawing share back from merchant GPUs; AMD started mass production of EPYC Venice on Zen 6 / 2nm. The "NVIDIA-only" narrative is starting to bend — hyperscaler in-house silicon is becoming material at the shipment level. (Tom's Hardware AI channel)
  • 05-31 | Microsoft × NVIDIA jointly announce Surface Laptop Ultra (powered by RTX Spark Superchip) — 20 ARM cores + Blackwell GPU + 128GB LPDDR5X unified memory, runs 120B-parameter models on-device, claims 1 petaflop AI compute. Announced on Computex opening day (T-1 week to GTC), positioned as the flare for "Windows on Arm, restarted." (Windows blog)
  • 05-31 | Jensen GTC Taipei keynote kicks off (PT 20:00 / Taipei 6/1 11:00) — the keynote has started, but specific announcements like Vera Rubin mass production, Cosmos 3, Nemotron 3 Ultra, and RTX Spark land per NVIDIA newsroom dates in W23. (keynote viewing guide)

🌍 Other vendor moves (aggregator view)

  • 05-27 | Fujitsu announces dual strategic partnerships with OpenAI and Anthropic the same day — Japan's largest SI firm bets on both frontier labs simultaneously. Newsletter editors rate it the classic sample of "enterprises can't pick just one anymore."9
  • 05-28 | CNN sues Perplexity AI for copyright infringement — CNN claims Perplexity scraped and redistributed its content without authorization. After NYT × OpenAI and Reuters × Cohere, another heavyweight name joins the AI-vs-media copyright docket; Perplexity's "Answer Engine" business model faces systemic compliance risk.10
  • 05-30 | Mystery enterprise burns $500M on Claude API in a single month — a consultant leaked that one unnamed company spent $500M on Claude calls in May alone — because they never set token caps on employee licenses + agentic tooling burns 1000× more. Tom's Hardware boosted it to viral over the weekend, making it the signature story of the "enterprise AI spend discipline era." (Tom's Hardware)
  • 05-31 | Reuters preview: Microsoft Build 2026 (6/2) will unveil the full in-house MAI lineup — the Mustafa Suleyman-led MAI division will drop MAI Image 2.5 / Voice 2 / Transcribe 1.5 / coding model / MAI-Thinking-1 reasoning model all at once, clearly aimed at wresting Copilot back from Claude Code. Detailed rollout lands next week in W23. (Tom's Guide)

Long reads

  • 05-26 | Jack Clark's Import AI 458 — Reckoning with the future — long essay + fiction (originally a 2026 Cosmos HAI Lab Oxford lecture), attempting a narrative for a "positive singularity." The most-shared AI long read on Substack notes this week.12
  • Backgrounder | Import AI 455 — Jack Clark estimates 60%+ probability of no-human-in-the-loop AI R&D by end of 2028; OpenAI's internal target for September 2026 is an "automated AI research intern." (source)

🧭 Trend analysis

1. Anthropic's valuation passes OpenAI for the first time — capital markets priced "who's ahead on the harness"

On 5/28 Anthropic's Series H closed at $965B post-money, leaping past OpenAI's $850B in one move. This is the first time in 2026 an AI company has clearly passed OpenAI on private-market valuation. The market's read isn't "Claude is that much stronger than GPT" (the model gap is actually small) — it's "Anthropic's harness engineering investments (Managed Agents / Self-hosted Sandbox / MCP Tunnels / Dynamic Workflows / Compliance API / Glasswing security automation) produce higher enterprise switching costs." What OpenAI needs to catch up on in H2 2026 isn't the model — it's the engineering.

2. Dynamic Workflows = the endgame form of harness engineering

The W21-W22 harness storyline (Smol AI's "harness is the new differentiator") finally got a phase-change sample on 5/28: one session orchestrating hundreds of parallel sub-agents to do repo-scale migration. This jumps straight from "PoC + manual handoff" to "fleet-scale, billable, replayable, monitorable." Impact on engineering teams: what used to be science fiction — "have AI refactor these 800 files for me" — is now a concrete workflow you can cut a PO for.

3. Hardware sits beside models as a co-headline for the first time

Lay this week's three big items side by side — Jensen's $150B Taiwan pledge, SK Hynix / Micron both crossing $1T, in-house ASIC shipments +44.6% YoY — and hardware / compute / memory enjoy "co-headline weight" with the model layer for the first time in 2026. The last two years were "win the model race → who buys more cards"; this year it's "who can get HBM, who can lock in TSMC capacity → who can train the model." This switch is structural for any enterprise CTO's 3-year capex plan.

4. "Enterprises can't pick just one" goes from intuition to canon

Fujitsu signing OpenAI + Anthropic on the same day is the cleanest sample, layered on top of W21's KPMG (Anthropic) + Dell (OpenAI). Enterprise AI procurement is migrating toward a multi-vendor mix of "two to three frontier + one to two domestic / open source". CIOs / CTOs need to add "single-vendor lock-in" to their risk models.

CNN × Perplexity (5/28) joins NYT × OpenAI and Reuters × Cohere on the copyright docket. Perplexity's "Answer Engine" model (scrape + rewrite + no traffic back) takes the most legal heat. Any AI product built on scraping for generation needs to re-examine its fair-use boundaries in H2 2026.

6. The Chinese camp this week was mostly "wrapping up + going quiet"

Kimi retired the K2 line on 5/25 and consolidated everything on K2.6; DeepSeek is in a silent sprint for V4.1 while pushing the domestic chip stack collaboration; Zhipu had no updates — all three are storing up powder for the next round (K3 / V4.1 / next-gen GLM). June is the key window to watch.

📚 Further reading