Week 23Published

Jensen's GTC Taipei triple-shot Vera Rubin/N1X/RTX Spark + Anthropic front-runs S-1 IPO + OpenAI lands on AWS Bedrock

nvidiagtc-taipeivera-rubinanthropicopenaiipocomputexqualcommintelmicrosofthbm
On this page

Date range: 2026-06-01 (Mon) → 2026-06-07 (Sun) Compiled: 2026-06-03 (Wednesday — covers the first 3 days; 6/4-7 weekend will be backfilled) Sources: NVIDIA Newsroom, Anthropic Newsroom, OpenAI, SiliconANGLE, Tom's Hardware, TechCrunch, Intel Newsroom, UPI, Blockonomi, and more.


🌐 Lead story (through 6-3)

"Hardware takes over the headlines + Anthropic front-runs the IPO + OpenAI breaks the channel lock" — three storylines in parallel.

W23 Monday through Wednesday is 2026's most concentrated hardware blitz —

  • 6/1 GTC Taipei keynote: Jensen refreshes NVIDIA's full stack in one shot in Taipei — Vera Rubin enters full production (40,000 engineers involved, billed as NV's largest project ever), the first consumer SoC N1/N1X Arm + Blackwell (128GB unified memory, RTX 5070-class performance), desktop superchip RTX Spark + Windows agentic OS, Foxconn upgraded to NCP self-building a 10,000 Blackwell AI Factory.
  • Same day 6/1, Anthropic files a confidential S-1 IPO draft with the SEC — just one week after closing the 5/28 Series H ($965B valuation). The first frontier AI lab to enter the SEC process; money, hiring, and model cadence across the entire ecosystem will move on this timeline.
  • Same day 6/1, OpenAI GPT-5.5 / 5.4 + Codex hit GA on AWS Bedrock — fulfilling the 4/28 strategic partnership. The "must go through Azure" era is over; Bedrock customers can call directly from commercial + GovCloud regions, billed at OpenAI prices but settled via AWS commit / IAM / PrivateLink / CloudTrail.
  • 6/2 Computex Day 1: Qualcomm declares "Year of the Agent" and unveils Dragonfly, its first datacenter AI chip (Qualcomm's first step into the inference chip market); Intel Crescent Island uses 480GB LPDDR5X to route around the HBM shortage; Samsung HBM5 prototype debuts globally with a 2nm base die; Micron HBM4 36GB 12H enters mass production; SK Group commits to doubling wafer capacity over 5 years — Jensen scrawled "Please Make More" on an HBM4E wafer at the SK hynix booth.
  • 6/2 Anthropic Project Glasswing scales to ~200 partners / 15+ countries (adds Samsung, NATO, and others); partners have already found >10,000 critical vulnerabilities; the "next few weeks" commit is to open Mythos-class models to all customers.
  • The Chinese camp stayed silent: DeepSeek V4.1 didn't ship in early June (was rumored for June); Kimi K3 likewise didn't ship. Both are still saving up for the next round.

⭐ 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 OpenAI GPT-5.5 / 5.4 + Codex GA on AWS Bedrock No more being forced onto Azure / no more direct-OpenAI lock-in — Bedrock customers can call immediately under AWS commit + IAM + PrivateLink, a switch-level change for heavy enterprise AWS users.2
🥈 2 Claude Code v2.1.160 security hardening Writing shell startup files (.zshenv/.bash_login) and build-tool config (.npmrc/.yarnrc/bunfig.toml/.bazelrc) now requires explicit confirmation — closing the prompt-injection-to-arbitrary-code-execution hole; Edit no longer requires a fresh Read after grep on a single file; WSL copy goes through PowerShell, clipboard fixed.11
🥉 3 Claude Platform advisor max_tokens + free streaming refusals The advisor tool adds a max_tokens cap, saving cost and cutting latency; when stop_reason: "refusal" fires with no Claude output, you are no longer billed. Two visible wins for engineers building guardrail routing. (docs)
4 NVIDIA RTX Spark Superchip + N1X laptop SoC (pre-order opening) Arm CPU + Blackwell GPU + 128GB unified memory makes running 120B / 300B models locally feasible; ships in fall on Surface Laptop Ultra, Dell XPS, ASUS ProArt. The second personal-AI-PC milestone after Apple Silicon.
5 AMD Ryzen AI Max Pro 400 The first x86 client CPU capable of running 300B-parameter models locally — another alternative to the NVIDIA path; worth a look for budget-constrained local-inference enthusiasts.

🏢 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 files confidential S-1 IPO draft with the SEC IPO process kicked off one week after the 5/28 Series H; $965B valuation → could overtake OpenAI's $850B once public. Every enterprise evaluating Anthropic needs to re-rank "long-term vendor sustainability" — public companies ≠ private companies on transparency, quarterly pressure, and product cadence.1
🥈 2 NVIDIA Vera Rubin enters full production + Foxconn upgraded to NCP Blackwell's successor Vera Rubin in full production (NV's largest project ever, 40k engineers); Foxconn upgrades from NV contract manufacturer to NV's own Cloud Partner (NCP), 10K Blackwell running in a Taiwan-domestic AI factory. A 3-year GPU roadmap + one more compute supplier on the menu. (NVIDIA News)
🥉 3 OpenAI flagship models + Codex GA on AWS Bedrock 4M weekly-active Codex + GPT-5.5/5.4 land on AWS commercial + GovCloud. Sovereign cloud / data-residency / FedRAMP customers can flip the switch now — the prior Azure lock-in is gone.2
4 Anthropic Project Glasswing scales to ~200 partners / 15+ countries + Mythos near-GA Samsung, NATO, and others added; partners have surfaced >10K critical vulnerabilities; "next few weeks" commit to open Mythos-class models to all customers. For power / water / healthcare / telecom / hardware critical-infrastructure enterprises, the leading defensive AI tool.12
5 Qualcomm Dragonfly datacenter AI chip + "Year of the Agent" Qualcomm steps into the datacenter for the first time and competes head-to-head with NVIDIA / AMD / Intel; same window also ships Snapdragon C (entry AI PC, $300 and up) + X2 Elite. AI inference chip suppliers go from 3 to 4 — one more notch of pricing leverage. (analysis)

🔵 Top two labs

Anthropic / Claude Code

Date Event One-liner
🔥 06-01 Confidential S-1 IPO draft filed Kicked off the SEC process one week after the 5/28 Series H. First frontier AI lab to enter the SEC process; ecosystem-wide money / hiring / model cadence will move on this timeline.1
06-01 Claude Code v2.1.160 (security hardening) Writing shell startup (.zshenv/.zlogin/.bash_login) and build-tool config (.npmrc/.yarnrc/bunfig.toml/.bazelrc, etc.) now requires explicit confirmation — closing the prompt-injection-to-arbitrary-code-execution hole; Edit no longer requires a fresh Read after grep on a single file; WSL copy goes through PowerShell, clipboard fixed.11
06-02 Claude Platform: advisor max_tokens + free streaming refusals advisor tool adds max_tokens cap to save cost and cut latency; when stop_reason: "refusal" fires with no Claude output, no billing. (release notes)
06-02 Project Glasswing scales to ~200 partners / 15+ countries + Mythos near-GA The first ~50 in April (Apple/Nvidia/Microsoft/CrowdStrike/Palo Alto) grow by +150 (incl. Samsung/NATO, covering power / water / healthcare / telecom / hardware across 15+ countries); >10K critical vulnerabilities surfaced; "next few weeks" commit to open Mythos-class models to all customers.12
06-03 AWS Summit Toronto keynote (livestream) The "first customer showcase after going on AWS," tied to the 6/1 Bedrock GA. (source)

OpenAI / Codex

Date Event One-liner
🔥 06-01 GPT-5.5 / 5.4 + Codex GA on AWS Bedrock Fulfilling the 4/28 strategic partnership. Bedrock customers can call directly from commercial + GovCloud regions — billed at OpenAI prices but settled via AWS commit, IAM, PrivateLink, CloudTrail. The "must go through Azure" era is truly over.2
06-01 Rosalind Biodefense — GPT-Rosalind early access Opens frontier biosciences reasoning model early access to trusted developers and U.S. federal / allied agencies; JHU APL does protein engineering, LLNL designs medical countermeasures. OpenAI's "biosafety" narrative response to Anthropic's Glasswing cadence.3

🟢 Chinese camp

DeepSeek / Kimi / GLM

  • Still silent this week. DeepSeek V4.1 (multimodal + MCP) was previously rumored for June; as of 6/3 there is no V4.1 entry on api-docs.deepseek.com/news/, likely delayed by compute / funding / talent constraints; Kimi K3 has its Polymarket / Manifold prediction-market window compressed to May-June (53% June 30), and Moonshot's official platform.moonshot.cn/docs/announcement has no update this week; Zhipu the same.
  • Observation: W22's Anthropic Series H + GTC Taipei pushed every Chinese vendor into internal "how do we respond" mode — the next release very likely carries an AVA Strategy framing (open-source pricing + compute-stack co-design + agent harness).

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

NVIDIA / Jensen Huang

Date Event One-liner
🔥 06-01 GTC Taipei keynote — Jensen declares "the agentic AI era has arrived" in Taipei Main stage 11:00 Taiwan time at the Taipei Music Center; a two-hour keynote refreshes the full stack: Vera Rubin / N1/N1X / RTX Spark / Cosmos 3 / DLSS 4.5 — "Taiwan's biggest product launch ever" (Jensen's own words). (NVIDIA keynote page, SiliconANGLE)
🔥 06-01 Vera Rubin enters full production Blackwell's successor; Jensen calls it "NV's most ambitious project ever, 40,000 engineers; 150 Taiwan supply-chain partners." HBM4 suppliers: Samsung + SK hynix + Micron all win the bid, SK hynix takes 60-70%. (Techtimes)
🔥 06-01 N1 / N1X Arm laptop SoC N1X = 20-core Arm CPU + 6144 CUDA cores (Blackwell GPU), up to 128GB LPDDR5X, 45-80W TDP; launch partners: Microsoft Surface, Dell XPS, ASUS ProArt, Lenovo. Performance lines up directly with the desktop RTX 5070 — Jensen formally steps into the Apple Silicon / Qualcomm X2 Elite ring. (videocardz)
🔥 06-01 RTX Spark Superchip — desktop / workstation Arm CPU + Blackwell GPU + 128GB unified memory; partners with Microsoft to turn Windows into an "agentic AI OS"; ships in fall. (Tom's Hardware)
🔥 06-01 Foxconn + Taiwan government + NVIDIA co-build AI Factory (10,000 Blackwell) Foxconn subsidiary Big Innovation Company becomes Taiwan's first NVIDIA Cloud Partner (NCP)Foxconn upgrades from "NV contract manufacturer" to "NV's own cloud partner." (NVIDIA News)
06-01 Rubin CPX — a new class of GPU for million-token context inference Integrated into the NVL144 CPX platform, positioned for million-token context inference; ships end of year. (newsroom)
06-01 Vera Rubin NVL72 with built-in Intelligent Power Smoothing Rack-level hardware mitigates the instantaneous grid shock from large-scale synchronous AI workloads. The "AI factory punches the grid" problem gets solved at the rack level — a new turning point in datacenter power design. (blogs.nvidia.com)

xAI / Tesla / Musk

  • No major news this week (W22 SpaceX S-1 aftershocks linger). Background: xAI Colossus 2 (Memphis) has expanded to 2GW / 555K GPU; Tesla Dojo 3 is exploring space-based AI compute, but no fresh first-party announcements on 6/1-6/3.

Other hardware (Intel / Qualcomm / AMD / Samsung / Micron / SK hynix / Foxconn)

Date Event One-liner
🔥 06-02 Qualcomm CEO Cristiano Amon announces "Year of the Agent" + Dragonfly datacenter AI chip Qualcomm enters the datacenter AI inference chip market for the first time — competing head-on with NVIDIA / AMD / Intel; same window ships Snapdragon C (entry AI PC, $300 and up) + X2 Elite + the Dragonwing IQ10 robotics reference design. AI inference chip suppliers go from 3 to 4. (analysis)
🔥 06-02 Intel Crescent Island AI inference GPU + Xeon 6+ (18A process) Crescent Island: Xe3P architecture, up to 480GB LPDDR5X (no HBM), 350W air-cooled; Intel differentiates on AI inference by "routing around the HBM shortage." Same day, Xeon 6+ goes online — Intel's first 18A-process datacenter CPU, 288 E-cores, paired with SambaNova SN-50 RDU for rackscale. (Tom's Hardware, Intel newsroom)
06-02 SK Group commits to doubling wafer capacity over 5 years + Jensen scrawls "Please Make More" SK Group chairman Chey Tae-won announced at Computex; Jensen wrote a hand-written plea-for-more on an HBM4E wafer at the SK hynix booth. The HBM crunch is tight enough that NVIDIA's CEO personally walks the memory-vendor booth. (UPI)
06-02 Samsung HBM5 prototype — world's first 2nm base die Samsung jumps HBM5 straight to a 2nm base — a technology-path overtake. (analysis)
06-02 Micron HBM4 36GB 12H module unveiled Claims 2.6x LLM inference throughput; stock rallied on the news. All three memory vendors now in line with HBM4. (Blockonomi)
06-02 AMD Ryzen AI Max Pro 400 Lisa Su unveils the first x86 client CPU capable of running 300B-parameter models locallyflanking play, doesn't fight NV for datacenter GPU, differentiates on "run 300B locally."
06-02 Foxconn global datacenter "token factory" strategy No longer just assembling NV machines — Foxconn now sells tokens itself. (Digitimes)
06-02/03 AMD Radeon RX 9070 GRE goes global Previously China-exclusive, $549, now globally available.
06-03 MediaTek cancels Computex keynote MediaTek's originally scheduled 6/3 keynote was pulled, handing over all narrative airtime for N1X (co-designed by MediaTek and NV) to Jensen.

Datacenters & power

  • Background: On 5/26 the Ministry of Economic Affairs confirmed an additional 5.2 GW of natural-gas capacity for the Taiwan grid before end of 2026 — a GTC Taipei precursor; AI compute expansion has pushed the Taiwan grid to the point where the government must officially add 5GW+.
  • 6/1 Vera Rubin NVL72's Intelligent Power Smoothing directly addresses the "AI factory punches the grid" problem (see the NVIDIA table above).

🌍 Other vendors (aggregator view)

  • Microsoft Build 2026 (6/2) prepares to launch the in-house MAI lineup (Reuters scooped it at the end of W22; Build officially opens this Wednesday) — MAI Image 2.5 / Voice 2 / Transcribe 1.5 / coding model / MAI-Thinking-1, led by Mustafa Suleyman, intent on pulling Copilot back from Claude Code's hands. Detailed rollout (full Microsoft Build keynote contents) will be backfilled into the second half of W23.

🧭 Trend analysis

1. GTC Taipei = NVIDIA enters a "full-stack refresh" cycle

Look at the 6/1 keynote in one sweep: Vera Rubin (datacenter) + N1/N1X (laptop) + RTX Spark (desktop / workstation) + Cosmos 3 (robotics) + Rubin CPX (million-token inference) — NVIDIA formally transitions from H100-single-point-GPU giant to a "datacenter ↔ desktop ↔ laptop ↔ robotics ↔ power" full-stack platform. For every company that depends on GPU, 3-year roadmaps need to be re-evaluated under the "NV full stack + Foxconn NCP + Windows agentic OS" assumption.

2. AI inference chip suppliers go from 3 to 4

W22 was "NVIDIA-only narrative loosening" (in-house ASICs at 27.8%); W23 is the formal expansion to 4: NVIDIA + AMD + Intel + Qualcomm. Qualcomm's Dragonfly is a brand-new tier of datacenter AI inference chip, plus Intel Crescent Island routes around the HBM shortage with LPDDR5X and AMD Ryzen AI Max Pro 400 covers x86 local 300B — buyer leverage at a historic high, but supply-chain complexity also expands.

3. HBM shortage formally enters the "CEO-personally-begs-for-more" phase

Jensen scrawling "Please Make More" on an HBM4E wafer at the SK hynix booth is the iconic moment of the week, but it's not PR — it's a real supply constraint: SK Group commits to doubling wafer capacity over 5 years + Samsung 2nm base die HBM5 prototype + Micron HBM4 36GB in production — all three vendors going flat-out still can't catch NV's Vera Rubin demand curve. That is also the root reason Intel differentiates on LPDDR5X.

4. The AI-company IPO era officially opens

Anthropic's 6/1 S-1 filing is the first by a frontier AI lab. Likely follow-ons: OpenAI advances its own IPO (previously rumored 12-18 months out) + Cohere / xAI / leading Chinese vendors (DeepSeek / Kimi) all reassess their capital structures. Impact on enterprise buyers: post-IPO companies differ from private ones on transparency, quarterly-report pressure, product cadence, and pricing stability — all of which must enter "long-term vendor sustainability" evaluation.

5. AWS channel break is OpenAI's strategic turning point

5 years ago OpenAI could only ship on Azure. Today GPT-5.5/5.4 + Codex are GA on AWS Bedrock. This is OpenAI's concrete move to "escape Microsoft lock-in" — combined with 4M weekly-active Codex developers, OpenAI is widening distribution from "enterprise LLM" into "cross-cloud developer platform." It also forms a direct channel-side standoff with Microsoft Build 2026's in-house MAI launch.

6. Anthropic doubles down on harness + security

Project Glasswing at ~200 partners + Mythos near-GA + Compliance API (W21) + Self-hosted Sandbox (W21) + Dynamic Workflows (W22) — Anthropic's strategic thesis is crystal clear: win against OpenAI's "widest distribution" via "highest enterprise switching cost." Which path is right will get an answer in the post-IPO earnings reports.

7. The Chinese camp's "no move" is itself a signal

DeepSeek V4.1 delayed + Kimi K3 delayed + Zhipu no update — three vendors collectively skip the early-June window, likely waiting for GTC Taipei + Anthropic IPO to settle before shipping. The window has shifted from "fight U.S./European vendors head-on on model cards" to "co-design the full stack once domestic chips mature." Late June to early July is the key observation period.


📚 Further reading


⚠️ Not yet covered this week

The items below are constrained by the fact that W23 still has four days (6/4-7) left to play out, and will be backfilled after Sunday:

  • 6/4 (Thu) Computex Day 3 + 6/5 (Fri) Computex Day 4 (Tesla / xAI / Chinese vendors may strike in this window)
  • Microsoft Build 2026 (from 6/2) full keynote contents — MAI lineup details
  • If the Chinese camp ships during 6/4-7 (DeepSeek V4.1 / Kimi K3 / Zhipu), it will be added here
  • WeChat public-account / X private-domain content (hard to auto-scrape outside Smol AI)