The #2 Cashes In

2026-09-28 — AMD agrees to buy Fei-Fei Li’s World Labs for $8.2 billion

AMD agreed to acquire Fei-Fei Li’s World Labs for about $8.2 billion in stock, with Li joining AMD as executive vice president and chief scientist AMD’s release says the deal is expected to close by the end of 2026, subject to regulatory approvals, and that World Labs’ spatial-intelligence models will inform AMD’s hardware and software roadmaps. CNBC notes AMD is chasing Nvidia, which already offers open-weight world models. (AMD, CNBC)

2026-09-22 — AMD crosses $1 trillion in market value, on a hyperscaler-customer product catalyst rather than its own numbers

AMD shares jumped 9.95% to close at $615.52 on Monday 09-21, taking its market capitalization above $1 trillion for the first time, on a rally triggered by Meta’s new consumer AI agent “Muse” hitting #1 on Apple’s App Store (448,000 daily active users ten days after launch) rather than by any AMD-specific news. Coverage frames the connection directly: AMD has a standing commercial relationship with Meta (the February 2026 agreement already on this thread, covering up to 6GW of AMD accelerator deployment), and traders read Muse’s rapid inference-traffic growth as demand validation for the CPU-plus-accelerator capacity AMD is building toward. It’s a new kind of catalyst for this thread to track — AMD’s valuation moving on a customer’s downstream product success rather than AMD’s own earnings, financing, or shipment news. (Bloomberg, Benzinga)

2026-09-11 — ARK keeps trimming AMD stock just ahead of the MI450 ramp

Cathie Wood’s ARK funds sold another ~$9.9m of AMD stock (19,491 shares via ARKK) on 2026-09-09, continuing a multi-week trim that predates AMD’s MI450/Helios accelerator ramp, even as coverage reports customer pull from Anthropic, OpenAI and Meta on MI450 as “very strong and tracking ahead of initial forecasts.” AMD remains ARK’s third-largest holding (~5.18% of net assets as of an April snapshot), so this reads as profit-taking into a >200% year-to-date rally rather than an exit — but it’s a second instance of ARK trimming a chip name (Nvidia earlier) just ahead of a hardware-revenue inflection, an early actual-money signal distinct from analyst commentary. (24/7 Wall St.)

2026-08-26 — A month after committing “tens of billions” to AMD, Anthropic signs a $45B Nvidia deal instead

Anthropic agreed to spend $45B over six years renting Nvidia Vera Rubin compute capacity from a planned 460MW Nscale campus in West Virginia (service starting late 2027) — reported 2026-08-26, roughly one month after Anthropic’s own up-to-2GW MI450 commitment to AMD (with AMD investing up to $5B into Anthropic on milestones). Coverage frames this as Anthropic refusing to let its AMD bet foreclose Nvidia’s next-gen platform, given Anthropic’s own $190-200B 2028 revenue projection. This is directly relevant to the thread’s underlying question of how durable AMD’s marquee customer commitments (OpenAI, Anthropic, Meta) actually are once a customer diversifies — Anthropic is the second-named customer in the thread’s title/entities to visibly hedge away from exclusivity. ⚠️ The Nscale deal is reported by multiple outlets citing sources, not yet confirmed by an Anthropic or Nscale press release found. (TechCrunch, CNBC, Yahoo Finance/Insider Monkey framing)

2026-08-24 — AMD lays out the actual MI400/Helios numbers at Hot Chips, past the roadmap slide it showed in June

AMD used its 2026-08-24 Hot Chips presentation to disclose hard specs for the Instinct MI455X GPU and the Helios rack it ships in, rather than the high-level roadmap language used at its June Investor Day: a 2.9-exaflop rack with 72 GPUs, 31TB of HBM4 at 1.7PB/s of bandwidth, and a single MI455X GPU rated at 432GB HBM4 / 23.3TB/s and 40.26 petaflops peak MXFP4 (4x the prior MI355X generation). The article does not give a firm shipping date beyond AMD’s own framing that Helios infrastructure is expected “to ship” to the hyperscaler partners it has been demonstrating the system to. This is the first hard-numbers disclosure since the thread’s last entry (the 2026-08-13 bond sale) and bears directly on whether AMD’s Instinct/Helios line can back up the demand its financing is now betting on. (ServeTheHome)

2026-08-13 — AMD prices its biggest-ever dollar bond sale to fund the buildout

AMD launched a four-part senior unsecured bond offering targeting $4-5B (notes due 2029/2031/2033/2036), ultimately pricing at $4.75B — its largest-ever US dollar debt raise. Proceeds are for general corporate purposes including possible debt repayment, as AMD funds its expansion into the AI/data-center compute this thread already tracks (the ~14GW of OpenAI/Meta/Anthropic deals, above). Bank of America, JPMorgan, Barclays and Wells Fargo led; bonds settle 08-17. First time this thread has seen AMD reach for public debt rather than cash/equity to fund the buildout — a leverage signal worth watching alongside the hardware-delivery timeline. (Bloomberg)

Watch: Whether the ~14GW of announced OpenAI/Meta/Anthropic compute deals actually convert to shipped hardware and vested warrants on their stated H2-2026/H1-2027 timelines, or stay a backlog/announcement story — tonight’s Q2 2026 earnings (due after this crawl closed) is the first real test. Also watching: China MI325X licensing status, and whether the Oracle 50,000-GPU Q3-2026 rollout starts on schedule.

2026-08-06→07 — AMD buys a model-specific-silicon startup

AMD acquired Taalas, a Toronto startup that hardwires specific AI models directly into silicon — its HC1 chip encodes Llama 3.1 8B across 53B transistors, quoted at ~17,000 tokens/sec/user at ~200W; Taalas says it can go from design to finished silicon in about two months. AMD plans to fold the approach into its Instinct accelerator roadmap as a differentiated-inference play; terms undisclosed. Announced 08-06, still leading AI-newsletter coverage into 08-07. (AMD Newsroom, CNBC)

2026-08-05 — Musk declares SpaceX/xAI “exclusive to Nvidia”; Wall Street re-rates AMD upward anyway

On SpaceX’s own earnings call, Elon Musk said SpaceX and xAI are now “exclusive to Nvidia” for future AI infrastructure — “we’ve decided to build exclusively on Nvidia because we think the Vera Rubin architecture is the best architecture” — tied to a joint SpaceX/Nvidia compute payload for the Starmind AI1 orbital AI satellite. He didn’t name AMD, but coverage reads it as closing off SpaceX/xAI as a future AMD customer, a competitive-narrative headwind landing the day after AMD’s Q2 beat. SpaceX itself reported $7.8B revenue (+92% YoY) but a $541M loss, and fell over 7% after-hours. (BNN Bloomberg · dnyuz)
Wall Street re-rated AMD upward despite the ~9% post-earnings drop, reading it as a valuation reset rather than a fundamentals miss: Truist raised its target from $478 to $594 (Buy), citing data-center revenue expected to double by 2027; Bernstein raised $600 to $650 (Outperform), citing the ~$13B Q3 guide beating consensus with data center projected at $8.1B vs. Street’s $7.7B; KeyBanc reaffirmed $725 (Overweight), citing AI GPU revenue of ~$2.9B, +190% YoY. (Finbold)

2026-08-04 — Q2 beats and raises; the stock falls anyway

Double beat, sold off ~8-9% after-hours on valuation/margin concerns. Revenue $11.536B (record, +50% YoY, above the $11.2B ± $300M guide); Data Center $6.718B (+107% YoY, above the ~$6.5B consensus); Q3 guide ~$13B. The Anthropic MI450/Helios partnership (up to 2GW) reaffirmed, first gigawatt now slated H1 2027 (not H2 2026 as earlier framed); Helios rack shipments named OpenAI, Meta, Microsoft, Oracle and Anthropic among recipients, shipments beginning later in Q3 2026. Coverage attributes the sell-off to stretched valuation (above Nvidia’s own multiples) and flat 56% margin guidance — Wall Street wanted a blowout Q3 forecast, not an in-line beat. (AMD IR)

← Backstory

Finding: artifacts/findings/amd-2026-08-04.md · bundle: artifacts/bundles/amd-2026-08-04/provenance.yaml

2026-08-04 — Q2 2026 earnings due tonight, pipeline still unproven in revenue

AMD reports Q2 2026 earnings today after market close (5:00pm ET call). AMD’s own guide from Q1: ~$11.2B revenue ± $300M; Street consensus ~$11.3B (+46-47% YoY), ~$1.61 non-GAAP EPS, Data Center segment estimated near $6.5B. As of this crawl, no 8-K with results had been filed on SEC EDGAR — the print itself falls after this crawl closed and is unresolved here.

2026-07-22/23 — Advancing AI 2026: MI400/Helios hit volume production, Anthropic signs

AMD’s Advancing AI 2026 event (San Francisco): Lisa Su presented the first volume-production details of the MI400 accelerator family and the Helios rack-scale platform (72x Instinct MI455X GPUs per rack, priced $5.0-5.5M/rack), plus 6th-gen EPYC Venice CPUs.
Same window, AMD and Anthropic announced a strategic partnership: Anthropic to deploy up to 2GW of AMD Instinct MI450 Series GPUs (first GW starting H1 2027), plus an AMD equity investment of up to $5B in Anthropic tied to deployment milestones, plus a joint ROCm/Claude engineering collaboration. Unlike the OpenAI/Meta deals, this is real equity, not a warrant — it already dominates AMD’s board-tracked thrust axis (~$5.3B/yr). Diversifies Anthropic’s compute supply beyond Google TPU/Cloud; gives AMD its second marquee frontier-lab customer after OpenAI. (AMD IR · CNBC)
Microsoft Azure confirmed the same event it will deploy Helios racks, with delivery expected before end of 2026 for frontier-model inference and Azure AI services. (CNBC)
Oracle Cloud committed to a 50,000-GPU rollout starting Q3 2026 (expansion of a partnership originally announced 2025-10-14); India’s Tata Consultancy Services also named as a deployment partner. Coverage described OpenAI’s + Meta’s commitments combined as 12GW, putting AMD’s total disclosed forward pipeline (with Anthropic’s 2GW) near 14GW.

2026-01 — China tightens: MI325X moves to case-by-case licensing

A new case-by-case US licensing regime began, requiring individual approval for MI325X shipments to Chinese buyers — tighter than the MI308 revenue-share arrangement below. Lisa Su has said China remains ~20% of AMD revenue despite the restrictions (single-source figure, not independently corroborated this crawl). (cryptobriefing.com)

2026-02-24 — Meta signs a second 6GW deal, days after expanding with Nvidia

AMD and Meta announced an expanded partnership: up to 6GW of AMD Instinct GPUs, days after Meta had also committed to millions of Nvidia GPUs — outlets framed it as Meta breaking Nvidia’s de facto exclusivity. First-gigawatt shipments (MI450-based GPU + EPYC Venice CPU, Helios architecture, ROCm) scheduled to begin H2 2026. AMD issued Meta a second performance-based warrant for up to 160 million shares, vesting on the same three-part structure as OpenAI’s (below): shipment tranches, AMD share-price thresholds, and Meta milestones. (AMD IR · CNBC)

2025-10-27 — ZT Systems manufacturing arm divested to Sanmina, design team kept

The Sanmina divestiture of ZT Systems’ US data-center manufacturing business closed (~$3B cash+stock incl. up to $450M contingent, announced May 2025) — AMD retained ZT’s rack-scale AI-solutions design and customer-enablement teams, the group doing the Helios integration work described above. (The underlying $4.9B ZT Systems acquisition itself had closed March 2025.) (AMD Newsroom)

2025-10-06 — OpenAI opens the playbook: 6GW, warrant to $600/share

AMD and OpenAI announced a strategic partnership: up to 6GW of AMD Instinct GPUs across multiple hardware generations, first 1GW (MI450-based) targeted for H2 2026. AMD stock jumped ~23% same day. AMD issued OpenAI a warrant for up to 160 million shares (~10% of shares outstanding), $0.01 exercise price, vesting on shipment tranches (1-6GW) plus AMD share-price thresholds plus OpenAI milestones — the final tranche requires AMD stock at $600/share; warrants run roughly five years, to October 2030. As of AMD’s FY2025 10-K/A (year ended Dec 27 2025), none of these warrant shares had vested — zero financial-statement impact through year-end 2025. (CNBC · Tom’s Hardware · SEC EDGAR — 10-K/A)

AMD disclosed an estimated ~$800M revenue hit from US export restrictions on the China-specific MI308 accelerator. MI308 exports then resumed under a conditions-based arrangement requiring AMD to share China-sales revenue with the US government — the same mechanism later reported for Nvidia’s H20. (qz.com · ainvest.com)