H100$6.39/hr 1.2% 7d
A100 80GB$2.45/hr 0.5% 7d
H200$10.29/hr 0.8% 7d
L40S$1.28/hr 0.3% 7d
T4$0.24/hr 0.6% 7d
L4$0.45/hr 1.1% 7d
H100$6.39/hr 1.2% 7d
A100 80GB$2.45/hr 0.5% 7d
H200$10.29/hr 0.8% 7d
L40S$1.28/hr 0.3% 7d
T4$0.24/hr 0.6% 7d
L4$0.45/hr 1.1% 7d
Company Analysis

Anthropic

Anthropic appears in Signwl's briefs as the principal compute-buyer signal in the inference-side of the custom-silicon substitution arc, with the $21bn TPU order (Broadcom / Google, 2026–2027 delivery) as the single largest non-GPU inference infrastructure commitment yet disclosed.

AI labs / frontier model developer (private)·Updated May 19, 2026

Get the Weekly Pulse — free

Signwl tracks Anthropic weekly. Subscribe for the next synthesis and the underlying pricing tape — Tuesday delivery, no spam.

Current Read

Anthropic has been a recurring reference across our recent briefs, appearing in 15 paragraphs across 4 articles. The thread is consistent: Anthropic is one of the most visible buyers in the inference-side of the custom-silicon substitution arc, with one specific deal (the $21bn TPU order) anchoring the narrative.

The $21bn TPU order. Per the May 17 brief: "Anthropic's $21B TPU order (Broadcom/Google, 2026–2027 delivery)." This is the largest single non-GPU inference infrastructure commitment in the publicly disclosed set. It validates the broader custom-silicon substitution thesis: a frontier model developer (Anthropic) is willing to commit balance-sheet-scale capital to non-Nvidia silicon, with Broadcom as the design-and-packaging partner and Google as the platform / chip-fab operator.

The timing nuance. The 2026–2027 delivery window matters. Per the same brief: "Inference shift to ASICs is real but lagging — the $21B Anthropic/TPU order is 2026–2027 delivery, meaning GPU inference is still the marginal capacity for 12–18 months." Anthropic's ongoing inference demand will continue to flow through Nvidia GPU rentals for the next 12–18 months. The TPU substitution is back-end-loaded.

The Claude Mythos / ExploitBench data point. Per the May 17 brief: "Claude Mythos is running ExploitBench at $36K per 122 episodes — these are GPU-intensive workloads that don't run on ASICs today." This quantifies one specific Anthropic workload's GPU intensity and clarifies that not all of Anthropic's compute is migrate-able to TPUs. Frontier training and high-complexity agentic workloads remain GPU-anchored.

The DeepSeek V4 efficiency context. Per the same brief: DeepSeek V4 (1.6T parameter MoE, launched April 24) requires only 27% of V3's single-token FLOPs and runs natively on Huawei Ascend without Nvidia. Even Anthropic's training-side compute curve is being recalibrated by upstream efficiency gains. This is a structural deflation in per-token compute demand that affects all frontier model developers, Anthropic included.

The composite picture: a key signal-carrier for the custom-silicon substitution arc, with one specific large deal that has been heavily quoted in our briefs and a back-end-loaded delivery profile that leaves GPU demand intact in the near term. Coverage is concentrated on the deal mechanics and the implications for the broader compute-buyer landscape, not on Anthropic's product trajectory.

Key Data Points

SignalSourceDate
Anthropic $21bn TPU order (Broadcom / Google) — largest single non-GPU inference commitment disclosedNews (May 12)2026-05-17
Delivery: 2026–2027 — GPU inference remains marginal capacity for ~12–18 monthsBrief synthesis2026-05-17
Claude Mythos running ExploitBench at $36K per 122 episodes — GPU-intensive workloadBrief synthesis2026-05-17
Named alongside OpenAI, Meta, Google as Broadcom XPU partnerNews (AOL May 12)2026-05-13

What to Monitor

  • TPU delivery progress. Updates to the 2026–2027 window — particularly any pull-forward — would tighten the substitution timing and worsen GPU demand expectations.
  • Anthropic GPU rental commentary. Anthropic doesn't disclose specific rental partner data, but any commentary on inference cost trajectory implicitly signals the substitution pace.
  • Claude pricing changes. Material reductions in Claude inference pricing would imply either successful ASIC substitution or per-token cost deflation (or both).
  • Custom-silicon partner expansions. Any new Anthropic chip-design partner announcement (beyond Broadcom / Google) would significantly extend the substitution thesis.

Recent Mentions

The Weekly Pulse

Track Anthropic and the broader AI infrastructure market.

A weekly synthesis of investable hypotheses and the underlying pricing tape — derived from Signwl's proprietary GPU pricing data. No fluff, no spam, unsubscribe anytime.

Disclaimer

The analysis on this page is synthesised from Signwl's published research briefs and is provided for general informational purposes only. It does not constitute investment, financial, legal, tax, or other professional advice. Signwl is not a registered investment adviser. Nothing on this page is a recommendation to buy, sell, or hold any security or financial instrument. Past performance does not guarantee future results. Readers should conduct their own analysis or consult a qualified professional before making investment decisions. Signwl makes no representation regarding the accuracy or completeness of third-party data referenced. The views expressed are those of Signwl Research at the time of publication and are subject to change without notice.

Methodology

This page is updated weekly when the new Weekly Pulse is published. The narrative is synthesised from Signwl's daily investment briefs and weekly pulses over the trailing 4–8 weeks. Pricing data is drawn from Signwl's proprietary regional pricing tape, blended across spot, on-demand, and 1-year reserved tiers from the major cloud providers. Source references are linked in the Recent Mentions section above.