In-House Silicon

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)

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)

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; caught by the 08-07 finalize’s coverage critic — 3 of 4 benchmark newsletters carried it while this thread’s terms only swept the hyperscalers’ chips, which is exactly why the thread had gone quiet). First frontier-lab entrant on a thread that had been hyperscaler-only; term and entity added. (TechCrunch)

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)

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.

Thread opened (ben-steer).
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.

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)

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)

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)

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)

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)