Autoresearch: Independent corroboration that two of three frontier models train without CUDA (Gemini/TPU, Anthropic/Trainium; OpenAI/NVIDIA as of late 2026)

Generated by /autoresearch on 2026-09-19. Synthesized across 3 rounds from 15 successful web pages (1 failed), anchored by Grokipedia entries for CUDA, Tensor Processing Unit, and Anthropic. See Provenance. Treat as raw material — review before promoting into a project or thread. Context: vault/projects/stock-market

Scope: training-side chip mix for frontier labs, to strengthen or honestly weaken Step 3 of cuda-moat-erosion-to-nvda-rerate (currently partial, single-sourced to Cerebras CEO Andrew Feldman, May 2026 Odd Lots). SemiAnalysis AgentX (2026-08-24) vs Jalapeño InferenceX (2026-08-25) is an unmerged inference-bench pair — not collapsed here. NVDA Q2 FY2027 print is hardware demand, not CUDA adjudication. No buy/sell/size. No new AI-infra chain.

Summary

Feldman’s May 2026 line — Gemini trained on Google TPUs, Anthropic on AWS Trainium with “no CUDA,” only GPT still CUDA-trained, hence a “70%” share loss — is directionally supported by official Google, Anthropic, Amazon, and OpenAI pages, but the exclusive “no CUDA / two-of-three” wording does not survive primary-source check. Google’s own papers, Cloud blog, Cloud Next 2026 remarks, Ironwood post, and 2026 Gemini 3.x model cards say Gemini is trained (and served) on TPUs, with JAX + Pathways named as the software — not CUDA. Anthropic and Amazon independently say Claude is trained and served on more than one million Trainium2 chips, with AWS as primary training and cloud partner (Project Rainier). Anthropic’s own 2025–2026 posts also say they train and run Claude on three platforms: Trainium, Google TPUs, and NVIDIA GPUs, plus a Nov 2025 NVIDIA architecture partnership (up to 1 GW Grace Blackwell / Vera Rubin) and a May 2026 SpaceX Colossus deal (>220,000 NVIDIA GPUs framed as Pro/Max capacity). OpenAI’s latest named frontier pretrain, GPT-6 Astra (Sep 2026), is described by OpenAI VP of Research Aidan Clark as the first pretrain on more than 100,000 GPUs at Stargate Texas; President Greg Brockman calls NVIDIA the preferred compute partner and “not changing”; OpenAI’s Sep 2025 letter of intent is for ≥10 GW of NVIDIA systems “to train and run” next-generation models. OpenAI’s Broadcom Jalapeño chip is an inference accelerator, not a training replacement. AMD Helios (MI455X) is a planned H2 2026 deploy under a 6 GW partnership, not the named Astra pretrain substrate. SemiAnalysis (28 Nov 2025) is the strongest non-Feldman, non-issuer restatement: Gemini 3 “trained entirely on TPUs”; Claude 4.5 Opus and Gemini 3 have “the majority” of training and inference infrastructure on TPUs and Trainium; Opus 4.5 trained on TPU among other hardware. “70% share” remains competitor framing, not an audited figure. Step 3 stays partial: Gemini/TPU and Anthropic-Trainium-primary are no longer Feldman-only; Anthropic “no CUDA” exclusive is weakened; OpenAI is still NVIDIA-pretrained on the latest named run.

Findings

Gemini: official Google sources say training is on TPUs (JAX/Pathways), not NVIDIA CUDA

Google’s December 2023 Cloud TPU v5p announcement states, of Gemini: “was trained on, and is served, using TPUs” (Google Cloud Blog, 2023-12-06). The same post quotes Jeff Dean: “TPUs are vital to enabling our largest-scale research and engineering efforts on cutting edge models like Gemini” (Google Cloud Blog, 2023-12-06).

The Gemini 1.0 technical report: “We trained Gemini models using TPUv5e and TPUv4 (Jouppi et al., 2023), depending on their sizes and configuration. Training Gemini Ultra used a large fleet of TPUv4 accelerators owned by Google across multiple datacenters” (arXiv 2312.11805). The same paper describes JAX + Pathways as the programming model and XLA/GSPMD as the partitioner — a Google compiler stack, not CUDA (arXiv 2312.11805).

The Gemini 2.5 report is more specific and later: “This model family is the first to be trained on TPUv5p architecture. We employed synchronous data-parallel training to parallelise over multiple 8960-chip pods of Google’s TPUv5p accelerators, distributed across multiple datacenters.” During that run, “93.4% of the time was spent performing TPU computations” (arXiv 2507.06261v2).

That TPU-training line is still live in 2026 product communications, not only 2023 papers:

• Amin Vahdat, Google Cloud Next 2026 8th-gen TPU post: “TPUs have been powering leading foundation models, including Gemini, for years.” TPU 8t is the training SKU (up to 9,600 chips / 2 PB HBM per superpod); TPU 8i is the inference SKU (blog.google, eighth-generation TPU).
• Sundar Pichai, Cloud Next 2026: “our TPUs, which have been so important in training and powering our Gemini models” (blog.google, Cloud Next 2026). The same post still offers “a portfolio of NVIDIA GPU instances” to Cloud customers — that is merchant GPU inventory, not a statement that Gemini pretrain moved onto CUDA (blog.google, Cloud Next 2026).
• Ironwood (TPUv7) post, 9 Apr 2025: “Leading thinking models like Gemini 2.5 and the Nobel Prize winning AlphaFold all run on TPUs today” (blog.google, Ironwood). “Run on” is serving + training color; it does not by itself prove a 2026 Gemini pretrain generation, but it is Google-primary and consistent with the papers.
• Gemini 3.1 Flash-Lite model card (published 3 Mar 2026): “Gemini 3.1 Flash-Lite was trained using Google’s Tensor Processing Units (TPUs).” Software: “Training was done using JAX and ML Pathways.” The card is based on Gemini 3 Pro (deepmind.google model card).

CUDA adjudication for Gemini: No fetched Google primary says “we do not use CUDA.” The affirmative claim is TPU hardware + JAX/Pathways. That is a non-CUDA stack as a matter of architecture (XLA/TPU, not nvcc/cuDNN). Treat “Gemini trains without CUDA” as supported inference from official hardware+software lines, not a Google sentence that uses the word CUDA.

Independent (non-Google, non-Feldman) restatement: SemiAnalysis, 28 Nov 2025: “Gemini 3 is one of the best models in the world and was trained entirely on TPUs.” Same piece: “SOTA models Gemini 3 and Opus 4.5 trained on TPU” (SemiAnalysis, TPUv7). Analyst, not issuer; still not an audit of FLOP-hours.

No fetched Google primary for 2025–2026 Gemini pretrain names NVIDIA GPUs as the training substrate. NVIDIA’s own Google-partnership blog discusses serving Gemini on NVIDIA GPUs in Vertex AI / Google Distributed Cloud (inference path) — keep that split (NVIDIA blog, Blackwell and Gemini was search-only in this pass, not fetched; do not load-bear on it).

Anthropic: Trainium is the stated primary training partner; “no CUDA” exclusive is not what Anthropic writes

Trainium / AWS — official, load-bearing, independent of Feldman.

Anthropic, 20 Apr 2026: “we currently use over one million Trainium2 chips to train and serve Claude.” Capacity under the new Amazon agreement: “up to 5 gigawatts (GW) of capacity for training and deploying Claude,” including “new Trainium2 capacity coming online in the first half of this year and nearly 1GW total of Trainium2 and Trainium3 capacity coming online by the end of 2026.” “We continue to choose AWS as our primary training and cloud provider for mission-critical workloads.” Andy Jassy (quoted on the same page): “Anthropic's commitment to run its large language models on AWS Trainium for the next decade” (anthropic.com, Amazon compute).

Amazon’s Project Rainier page: cluster of “nearly half a million Trainium2 chips”; “Anthropic is actively using Project Rainier to build and deploy its industry-leading AI model, Claude”; “Claude is expected to be on more than 1 million Trainium2 chips—for workloads including training and inference—by the end of the year”; Rainier is “more than five times the compute power Anthropic used to train its previous AI models” (aboutamazon.com, Project Rainier).

AWS Trainium customers page, attributed to James Bradbury, Head of Compute, Anthropic: “With almost a million Trainium2 chips training and serving Claude today, we're excited about Trainium3…” (aws.amazon.com, Trainium customers).

TPU is also in Anthropic’s training mix — also non-CUDA, which cuts both ways for Feldman.

Anthropic, 23 Oct 2025: expansion “including up to one million TPUs,” “worth tens of billions of dollars,” “well over a gigawatt of capacity online in 2026.” Then the sentence that breaks the exclusive Trainium/no-CUDA line: “Anthropic’s unique compute strategy focuses on a diversified approach that efficiently uses three chip platforms–Google’s TPUs, Amazon’s Trainium, and NVIDIA’s GPUs.” Same post: “We remain committed to our partnership with Amazon, our primary training partner and cloud provider, and continue to work with the company on Project Rainier” (anthropic.com, expanding TPU use).

Anthropic, 6 Apr 2026 (Google + Broadcom, multiple GW of next-gen TPU from 2027): “This significant expansion of our compute infrastructure will power our frontier Claude models.” And again: “We train and run Claude on a range of AI hardware—AWS Trainium, Google TPUs, and NVIDIA GPUs—which means we can match workloads to the chips best suited for them. … Amazon remains our primary cloud provider and training partner” (anthropic.com, Google–Broadcom).

NVIDIA is in the mix. Feldman “no CUDA” is not Anthropic’s wording.

Anthropic, 18 Nov 2025, Microsoft + NVIDIA: “Anthropic to adopt NVIDIA architecture.” “Anthropic’s compute commitment will initially be up to one gigawatt of compute capacity with NVIDIA Grace Blackwell and Vera Rubin systems.” Closing line: “Amazon remains Anthropic’s primary cloud provider and training partner” (anthropic.com, Microsoft–NVIDIA). The post does not say that 1 GW is pretrain vs serve. Do not invent the split.

Anthropic, 6 May 2026, SpaceX Colossus 1: “more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month. This additional capacity will directly improve capacity for Claude Pro and Claude Max subscribers.” That sentence is subscriber/serving capacity, not a named pretrain. The three-platform line is repeated (anthropic.com, SpaceX).

Independent analyst (not Feldman, not issuer): SemiAnalysis, 28 Nov 2025: “The two best models in the world, Anthropic’s Claude 4.5 Opus and Google’s Gemini 3 have the majority of their training and inference infrastructure on Google’s TPUs and Amazon’s Trainium.” Separately: “Anthropic training Sonnet and Opus 4.5 on multiple types of hardware including TPUs.” Paywalled remainder not used (SemiAnalysis, TPUv7).

How to read this against Step 3. Anthropic does train Claude at frontier scale on non-CUDA silicon (Trainium as primary partner; TPU as a second large non-CUDA fleet). That strengthens the “Anthropic is not CUDA-only / not NVIDIA-only at training” half of Feldman. Anthropic does not claim “no CUDA.” They name NVIDIA GPUs in the train-and-run set and signed a GW-scale NVIDIA architecture deal. Without an audited FLOP-hour split, “Anthropic models trained on Trainium, no CUDA” is too strong. Honest formulation: primary training partner = AWS/Trainium; material TPU training; NVIDIA present, split unaudited, some of the named NVIDIA capacity is serving (SpaceX Pro/Max).

OpenAI / GPT: latest named frontier pretrain is NVIDIA GPUs; custom silicon on the board is inference; AMD is planned, not Astra

NVIDIA training — OpenAI-primary and NVIDIA-primary, mid-to-late 2026.

OpenAI and NVIDIA, 22 Sep 2025 letter of intent: “at least 10 gigawatts of NVIDIA systems for OpenAI’s next-generation AI infrastructure to train and run its next generation of models.” “OpenAI will work with NVIDIA as a preferred strategic compute and networking partner.” First GW targeted H2 2026 on Vera Rubin. Greg Brockman on that page: “We’ve utilized their platform to create AI systems that hundreds of millions of people use every day” (openai.com, NVIDIA 10 GW).

Greg Brockman, Stratechery interview around Astra: “this is the first run that we’ve trained on more than 100,000 GPUs.” On Jalapeño vs NVIDIA: “We partner very closely with Nvidia as our preferred compute partner, and if you look at the size of the computers we’re building and the unique computers we’re building, we need Nvidia, there’s no question about it, we’re building these amazing training computers, we’re building lots of inference with them.” Later: “Nvidia is our preferred partner, that’s not changing. In fact, we’re leaning in even more with them” (Stratechery, Brockman on Astra).

OpenAI VP of Research Aidan Clark, to reporters at the Astra launch, quoted by Fortune (3 Sep 2026): “It’s the first time we’ve pretrained on more than 100,000 GPUs at our Stargate site in Texas.” Fortune also attributes “largest training run by far” to Clark (Fortune, GPT-6 Astra). That is OpenAI research leadership, not Feldman, not Huang’s X post. (Huang’s “~100K+ Nvidia Grace Blackwell NVLink72” line was not fetched as a hydrated X permalink in this pass; do not treat PC Gamer recaps as primary.)

Jalapeño is inference, not a training-chip substitution. OpenAI, 24 Jun 2026: “Jalapeño, OpenAI’s first Intelligence Processor: an accelerator architected around OpenAI’s vision for the future of LLM inference.” Richard Ho: “Jalapeño was designed from the ground up for LLM inference” (openai.com, Broadcom Jalapeño). The follow-up results page (/jalapeno-first-results/) 403’d this pass; search snippets attributed a line that OpenAI will “continue to widely deploy accelerators from NVIDIA and other partners for both training and inference” — not used as a citation because the page was not fetched. SemiAnalysis’ InferenceX Jalapeño write-up is the unmerged inference bench already on the mechanism page; it does not adjudicate training CUDA (SemiAnalysis Jalapeño — search dump only; paywall after the CUDA-moat-at-inference aside).

AMD is a second GPU vendor in the OpenAI plan, not evidence Astra trained off NVIDIA. AMD newsroom, 23 Jul 2026: collaboration began with MI300X, deepened with MI355X; Helios (MI455X) access “for several months”; “The companies are optimizing GPT-class workloads on the platform”; OpenAI “expects to begin bringing Helios online through multiple deployment partners in the second half of 2026,” first phase of the Oct 2025 6 GW deal (AMD newsroom, AAI 2026). That is ROCm-path GPU compute in the future mix, not a statement that GPT-6 Astra was AMD-pretrained. If Helios ramps, Feldman’s “only GPT remains CUDA-trained” would need a 2027 revisit (AMD is not CUDA). As of 2026-09-19, the named Astra pretrain is GPUs at Stargate, with NVIDIA as preferred partner.

SemiAnalysis (Nov 2025) said OpenAI “hasn’t even deployed TPUs yet” in that piece’s TCO aside (SemiAnalysis, TPUv7). That is dated relative to Sep 2026 Astra; it is not a 2026-09 OpenAI TPU-training confirmation. Do not upgrade it.

CUDA vs “NVIDIA GPUs” for OpenAI. OpenAI’s fetched pages say NVIDIA systems / GPUs / preferred compute partner. They do not say “we train on CUDA.” NVIDIA GPUs in production training almost always imply the CUDA stack; that is still an inference, same as TPU→not-CUDA. Do not pretend OpenAI issued a CUDA market-share number.

What this does to Feldman Step 3 (and what it does not)

Implication for cuda-moat-erosion-to-nvda-rerate Step 3: evidence quality upgrades (multiple issuer primaries + one independent analyst family). Step status should stay partial, not confirmed, because (a) Anthropic is multi-vendor including NVIDIA, (b) CUDA is rarely named, (c) 70%/2-of-3 is still not a measured share, (d) OpenAI AMD deploy could change the “only GPT” clause after Helios ramps. Do not mint a new AI-infra chain. Do not collapse AgentX vs Jalapeño. Do not re-date NVDA off this mix.

Grokipedia’s TPU and Anthropic pages already encoded the dual Anthropic TPU+Trainium partnership; they are encyclopedic primers, not substitutes for the 2026 issuer posts (Grokipedia TPU, Grokipedia Anthropic). Grokipedia CUDA is the 2006–2025 NVIDIA stack history, not a 2026 lab mix (Grokipedia CUDA).

Contradictions and open questions

• Anthropic NVIDIA FLOP-hours vs Trainium/TPU FLOP-hours at pretrain — unaudited. Official language supports Trainium as primary partner and NVIDIA as in the mix. SpaceX 220k GPUs are explicitly Pro/Max capacity. The 1 GW Grace Blackwell / Vera Rubin commitment does not say pretrain vs serve. SemiAnalysis “majority … on TPUs and Trainium” is the only independent split-like claim; it is not a table of exaFLOP-days.
• “Without CUDA” vs “not NVIDIA silicon.” TPU = JAX/Pathways/XLA (named). Trainium = AWS Neuron (not fetched as a software-stack primary this pass; do not over-claim Neuron details). NVIDIA GPUs ≠ a published “we use CUDA” sentence. Feldman used CUDA; issuers use chip names. A wiki ingest should not silently equate them without a qualifier.
• Gemini 3 Pro PDF model card (storage.googleapis.com) was not fetched (PDF off the *.gov whitelist). HTML 3.1 Flash-Lite card (based on 3 Pro) was fetched and names TPUs + JAX/Pathways. Generation/SKU of TPU for Gemini 3 (v5p vs Ironwood vs 8t) is not in the fetched HTML card. LinkedIn/intuitionlabs claims of “v5e and v6e pods only” were not treated as primary.
• OpenAI AMD MI300X/MI355X “previously deployed” (DCD/search; AMD newsroom confirms the collaboration path) — not shown as the Astra 100k-GPU pretrain. Whether any GPT-class pretrain has completed on ROCm remains open. Helios H2 2026 is the date to watch for a Step 3 revisit.
• OpenAI TPU adoption — SemiAnalysis Nov 2025 said not yet deployed; no 2026 OpenAI primary fetched here that names TPU training. Gap, not a negative proof.
• Meta / xAI / others — out of the “three leading models” frame; not researched. Do not extend 2-of-3 beyond Gemini / Claude / GPT as Feldman used it.
• AgentX vs Jalapeño InferenceX — still unmerged, still inference. Jalapeño’s existence does not move Step 3. SemiAnalysis’ “CUDA moat is potentially dead” line in the Jalapeño piece is an inference-software aside, already on the mechanism page as inference-side color.
• NVDA Q2 FY2027 — hardware demand; not used.

Provenance

Rounds run: 3 of 3 (full). No early exit: Anthropic NVIDIA mix and OpenAI AMD/Jalapeño were material to whether Step 3 can leave partial.

Sub-questions by round:

Round 1 (broad survey):

1. Does Google officially document Gemini training on TPUs?
2. Does Anthropic/AWS officially document Claude training on Trainium?
3. Does OpenAI still train GPT primarily on NVIDIA GPUs as of mid-to-late 2026?
4. Do issuer pages describe a chip mix rather than exclusive stacks?
5. Is there a non-Feldman restatement of “two of three”?

Round 2 (drill-down):

1. Anthropic NVIDIA: training or serving? — targeting Feldman “no CUDA.”
2. Gemini 3.x / 2026 still TPU-trained? — targeting post-2.5 continuity.
3. OpenAI-named GPU counts for Astra — targeting a non-Huang primary.
4. SemiAnalysis training-side TPU/Trainium — targeting independent analyst corroboration.

Round 3 (resolve remaining uncertainty):

1. Fetchable HTML Gemini 3.x model-card hardware line (not off-whitelist PDF).
2. AMD Helios vs Astra pretrain substrate.
3. SpaceX NVIDIA GPUs: serving vs training.
4. Ironwood “Gemini 2.5 runs on TPUs” as 2025 Google-primary continuity.

Anchor source (Grokipedia, fetched before round 1):

• CUDA — 12,017 chars excerpted (API) — NVIDIA software-stack primer; no 2026 lab mix.
• Tensor Processing Unit — 10,017 chars excerpted — Gemini/TPU history; flags Anthropic’s Oct 2025 TPU expansion for training and serving.
• Anthropic — 8,017 chars excerpted — dual AWS Trainium and Google TPU partnerships already in the encyclopedic record.

URLs fetched (15 successful, 1 failed):

Round 1:

• Expanding our use of Google Cloud TPUs and Services (Anthropic, 2025-10-23) — official — three-platform mix; up to 1M TPUs; Amazon primary training partner.
• Anthropic and Amazon expand collaboration for up to 5 GW (2026-04-20) — official — >1M Trainium2 to train and serve Claude; AWS primary training provider.
• Anthropic–Google–Broadcom next-gen TPU (2026-04-06) — official — train-and-run on Trainium, TPU, and NVIDIA GPUs.
• Eighth generation TPUs: two chips for the agentic era (Google, Cloud Next 2026) — official — TPUs powering Gemini for years; 8t training / 8i inference.
• Introducing Cloud TPU v5p (Google Cloud, 2023-12-06) — official — Gemini trained and served on TPUs; Dean quote.
• AWS activates Project Rainier — official Amazon — ~500k Trainium2; Claude training and inference toward >1M chips.
• OpenAI and Broadcom unveil Jalapeño inference chip (2026-06-24) — official — Jalapeño is LLM inference, not training.
• [Failed: https://openai.com/index/jalapeno-first-results/] — 403 Forbidden. Search snippets not treated as fetched.
• Gemini 2.5 paper arXiv 2507.06261v2 — academic HTML — first family on TPUv5p; 8960-chip pods; 93.4% TPU compute time (full-text via search extract).
• Gemini 1.0 paper arXiv 2312.11805v5 — academic HTML — TPUv4/v5e; Ultra on TPUv4 SuperPods.

Round 2:

• Microsoft, NVIDIA, and Anthropic partnerships (2025-11-18) — official — up to 1 GW Grace Blackwell / Vera Rubin; Amazon still primary training partner.
• OpenAI–NVIDIA 10 GW systems partnership (2025-09-22) — official — train and run next-gen models on NVIDIA; preferred partner; first GW H2 2026 Rubin.
• AWS Trainium customers — official — Bradbury: almost a million Trainium2 training and serving Claude.
• Fortune: GPT-6 Astra launch (2026-09-03) — news, quotes OpenAI VP Research — first pretrain on >100,000 GPUs at Stargate Texas.
• Stratechery: Brockman on Astra — interview extract — >100k GPUs; NVIDIA preferred partner “not changing.”
• SemiAnalysis: TPUv7 / 900lb gorilla (2025-11-28) — independent analyst, free lede — Gemini 3 trained entirely on TPUs; majority of Opus 4.5 + Gemini 3 infra on TPU/Trainium; Opus 4.5 trained on TPU among other hardware. Paywall not used.
• Cloud Next 2026, Pichai — official — TPUs important in training and powering Gemini.

Round 3:

• Gemini 3.1 Flash-Lite model card (2026-03-03) — official DeepMind — trained on TPUs; JAX and ML Pathways.
• AMD and OpenAI: AAI 2026 update (2026-07-23) — official AMD — Helios H2 2026 start of 6 GW; GPT-class workload optimization; not Astra pretrain.
• Anthropic: SpaceX Colossus / higher limits (2026-05-06) — official — >220k NVIDIA GPUs for Pro/Max capacity; three-platform line repeated.
• Ironwood TPU blog (2025-04-09) — official — Gemini 2.5 and AlphaFold run on TPUs today.

Not fetched (reliability / policy): Gemini 3 Pro PDF model card on storage.googleapis.com (PDF off *.gov whitelist). SemiAnalysis Amazon Trainium expansion piece beyond the TPUv7 lede (paywall). Huang X post (no hydrated permalink this pass). NVIDIA Google-partnership blog (search-only).

Tools used: WebSearch, WebFetch, grokipedia.py (CUDA, Tensor Processing Unit, Anthropic). No --include-x. No --fanout. No --auto.

Generated: 2026-09-19 ~20:00 UTC