2026-10-20 — Google releases TorchTPU as open source, a day after Amazon’s TorchNeuron

• Google announced on Tuesday that TorchTPU, its version of PyTorch, is available as an open-source project, TechCrunch reported — the article says it lets AI apps run on Google’s Tensor Processing Units and lets existing PyTorch apps move to TPUs and Google Cloud with a configuration change and no rewriting. (TechCrunch) Claims: kat-claim-5ee1826be130
• Amazon released its own PyTorch version, TorchNeuron, which lets AI apps be ported to its Trainium chip and to AWS with no rewriting, TechCrunch reported — the article says Amazon announced it on Monday and is also submitting code to the main PyTorch project that decouples features from Nvidia’s chips. (TechCrunch) Claims: kat-claim-a27afd311125

2026-10-12 — Microsoft announces Maia 300, its third in-house AI chip, to be its main inference system from 2027

• Microsoft announced Maia 300, its third-generation in-house AI chip, designed to be the primary AI inference system for its products and services and to start rolling out to production workloads in 2027, The Verge reported — Microsoft’s Scott Guthrie said the company expects 40 percent more tokens per megawatt than leading GPU-based systems and will scale the Azure Maia family to multi-gigawatt deployments starting in 2027; The Verge reports the chip is built on TSMC’s N2P process with a chiplet design separating compute, memory controller and I/O. These are Microsoft’s own performance figures, not independent measurements. (The Verge) Claims: kat-claim-04da171bc68f, kat-claim-1604732130d2, kat-claim-8ddbe389ce6a, kat-claim-2c38b20a50a0, kat-claim-7c004023627f

2026-09-30 — AWS signs a $1 billion-plus chip-design licence with Synopsys

• Amazon Web Services agreed a multi-year deal worth more than $1 billion to license chip-design intellectual property from Synopsys, which names Amazon the lead customer for its new “application-optimized” silicon IP The deal moves Synopsys to a license-plus-royalty model with payments tied to production volumes, extends AWS’s use of Synopsys’s design software and agentic AI tools, and does not say which AWS chips will use the designs. (Synopsys, Quartz via Yahoo Finance)

2026-09-22 — Google pitches TPUs to Korea, outside its own cloud

• Google has offered South Korea priority supply of its TPU accelerators for the country’s sovereign-AI data-center buildout, following an April announcement that it would sell TPUs to data centers outside Google Cloud. Seoul Economic Daily reports Google met Korean AI-strategy officials on 09-09, cites Morgan Stanley’s estimate of Google direct TPU sales of $84bn in 2027 and $108bn in 2028, and notes Anthropic’s plan to expand TPU use from 1GW this year to 5GW next year; an unnamed Korean industry official told the paper Nvidia GPU delivery lead times run “as long as 30 weeks” and that Google is exploiting the gap. It moves the TPU from an in-house inference chip toward a product sold to third parties and governments, which is a step past the inference-only framing on this thread. (Seoul Economic Daily)

2026-09-17 — Broadcom’s Sept 2 earnings call named OpenAI as a 4th custom-silicon customer, $10B in orders, Jalapeño already shipping

• Broadcom’s September 2 earnings call confirmed OpenAI as a named fourth AI-chip customer, alongside $10B in AI-chip orders and confirmation that the Jalapeño chip (already benchmarked at Hot Chips, 08-25 entry above) shipped this quarter, on track for a 1.3GW deployment. Caught on this thread’s cold-rotation pass, 15 days after the earnings call itself — a lag worth noting given how central Jalapeño’s production timeline is to this thread’s thesis. (Broadcom Q3 FY2026 earnings call, 2026-09-02)

• Qualcomm and Amazon announced a multi-generation collaboration on custom AI-inference silicon and optical interconnects for AWS data centers, with Qualcomm also issuing Amazon warrants for $4B of Qualcomm stock and AWS committing to purchase up to $60B of Qualcomm chips and technology over the deal’s life. This is a new fork in the thread’s thesis: rather than one hyperscaler-designed chip line per company, Amazon is now running Trainium in-house and buying custom AWS-specific silicon from a merchant vendor (Qualcomm, diversifying beyond its smartphone base) — a hybrid pattern this thread hasn’t tracked before, distinct from Broadcom/Marvell’s co-design-house role. Qualcomm’s own release names the technology (custom inference silicon, SerDes/optical DSP, 1.6T+ interconnects) but discloses no dollar figures; the $4B warrant and $60B purchase-commitment figures come from Qualcomm’s SEC-linked disclosure as reported by CNBC and Bloomberg. (Qualcomm/PR Newswire, CNBC)

2026-09-02 — Broadcom names Anthropic its likely largest custom-chip customer, with 5GW of next-generation TPUs planned for 2027

• Broadcom CEO Hock Tan said on the company’s 09-02 earnings call that Anthropic “is on track to become our largest XPU customer in 2027 and sustain that in 2028.” Broadcom is deploying 1 gigawatt of Google’s Ironwood TPU capacity for Anthropic in 2026, expects Anthropic to deploy another 5 gigawatts of the next-generation TPU v8i in 2027, and sees a further 10 gigawatts of potential Anthropic deployments in 2028; Tan also said production shipments of TPU v8i have begun. This is the sharpest evidence yet for this thread’s anchor claim that a frontier lab running on custom silicon (rather than Nvidia GPUs) is real: the customer Broadcom expects to pass Google itself is a tenant, not a chip designer. It sits beside the OpenAI fourth-customer entry logged 09-17. (Benzinga, SiliconANGLE)

2026-08-25 — Jalapeño’s first public benchmarks land at Hot Chips

• OpenAI and Broadcom published Jalapeño’s first performance benchmarks at Hot Chips 2026, claiming better performance-per-watt than Nvidia’s flagship. The inference ASIC the two announced in June was disclosed at 700W TDP, with a 128-chip pod at 1.7 exaflops in 4-bit and 27.5TB of HBM4, and claimed advantages up to 1.9x throughput-per-kilowatt and 3.6x lower latency on SemiAnalysis’s InferenceX benchmark; small-scale deployment inside OpenAI is targeted for end-2026, broader rollout 2027. The June item was a partnership announcement — this is the first evidence the silicon works. (OpenAI, TechCrunch)

2026-08-21 — Anthropic hires the founding architect of Google’s TPU program

• Amir Salek, who founded and led Google’s custom-silicon division from 2013 to 2022 across the first seven generations of the TPU, joined Anthropic’s compute team, reporting to James Bradbury, as part of a push to design its own AI chips. This thread has read the custom-silicon revolt as real but inference-only, anchored by Anthropic running on Trainium and TPU rather than building. This is the first hard evidence it intends to design — and the most senior custom-silicon hire any frontier lab has made. (Bloomberg, Crypto Briefing, Seeking Alpha)

2026-08-06 — Anthropic joins the in-house silicon story

• Anthropic confirmed it is building an in-house AI chip-design team, targeting ~50% inference-cost cuts, with Samsung scouted as a fab partner (disclosed 08-05; carried by 3 of 4 benchmark newsletters). First frontier-lab entrant on a thread that had been hyperscaler-only. (TechCrunch)

2026-08-05 — A fifth entrant: Anthropic starts building its own chip-design team

• Anthropic is assembling an internal AI chip-design team, confirmed by the company after Business Insider broke the story — recruiting through a “custom silicon team” job listing, aiming to “co-design hardware and models to help its technology run faster and more efficiently.” Anthropic currently relies entirely on external partners (AWS, Google, Nvidia, AMD) for compute; The Information separately reported it is evaluating Samsung as a manufacturing partner. This is the customer side of the story, not another hyperscaler program — the exact company this thread cites as the anchor case for “inference-only, still mostly Nvidia plus Trainium+TPU” is now moving toward designing silicon of its own rather than only consuming hyperscaler chips. Coverage doesn’t specify whether the target is training or inference/serving — a genuine open question, not yet resolved by any source. (TechCrunch)

2026-07-27 — Four programs, one wall

• TPU v7 Ironwood (Broadcom/MediaTek/TSMC co-design) · Maia 200 (dual-purpose, volumes undisclosed) · Trainium 1.4M deployed · MTIA “millions” H100-eq, Iris Sept-26. All four decline to state the own-vs-Nvidia split — the wall every capex crawl hit.

2026-07-24 — Opened + crawled

• Thread opened.
• Verdict: diversification at the margin and on inference, not a rout. The CUDA-vs-XLA software moat keeps Nvidia the default for anyone without a captive anchor customer; the “in-house” silicon is largely outsourced to Broadcom/Marvell. ⚠ WebSearch budget exhausted (fan-out) — primary-doc fetch path; Maia 200 + some codenames single-sourced, flagged.

← Backstory (crawl 2026-07-24 → artifacts/findings/inhouse-silicon-2026-07-24.md)

2026-06 — Broadcom is the real winner

• Broadcom’s custom-AI revenue hit $10.8B in a single quarter (+143% YoY), guiding $29.4B next and reaffirming “>$100B AI revenue in 2027”; ~6 hyperscaler XPU customers. Marvell (Trainium/Maia co-design) projecting ~$11B AI-ASIC revenue in 2026. The revolt enriches the co-designers. (Reuters)

2026-04 — Google Ironwood + the Anthropic anchor

• Ironwood (TPU v7) hit broad Cloud GA (04-22); combined with the Oct 2025 Anthropic–Google deal (up to 1M TPUs / >1GW in 2026, first phase 400k Ironwood). SemiAnalysis: Ironwood ~44% lower TCO than a GB200 server; Anthropic ~52% lower TCO/PFLOP vs GB300 — the economics behind the revolt. (SemiAnalysis)

2026-03 — Meta MTIA: inference, not frontier

• Meta’s MTIA 300 in production (ranking/rec), 400/450/500 on a 6-month cadence, “hundreds of thousands” deployed for inference — but explicitly NOT frontier-LLM training (Meta still buys Nvidia/AMD for that). The workload-specific pattern. (Meta)

2026-01 — Microsoft the laggard

• Maia 200 launched — inference-only (TSMC 3nm, claims 30% better perf/$); next-gen “Braga” reportedly delayed into 2026. Microsoft is furthest behind and still overwhelmingly Nvidia-reliant for training. ⚠ Maia 200 specs are a vendor claim, single-sourced — don’t treat as independent benchmark. (Microsoft)

2025-10→12 — The Anthropic-on-Amazon anchor

• Amazon Trainium3 UltraServers GA (re:Invent, up to 50% lower cost); Project Rainier live (~500k Trainium2 for Claude training, Indiana); AWS passed 1M+ Trainium deployed (+150% YoY); Anthropic committed >$100B to AWS/10yr, up to 5GW Trainium. A frontier lab running mostly off Nvidia is the strongest evidence the revolt is real. (Anthropic)