ai drama

Anthropic Is 'Exploring' Custom Chips. That Is the Most Expensive Way to Say No.

What launched / what broke

On April 10 Reuters reported via three sources that Anthropic is exploring its own AI chips, with plans that are pre-team and pre-commitment. Claude inference currently runs on a mix of Google-Broadcom TPUs, Amazon Inferentia-Trainium chips, and Nvidia GPUs. The company signed for 3.5 GW of TPU capacity starting 2027, three times its 2026 usage and contingent on continued growth, while the Broadcom-Google TPU supply agreement extends through 2031 with one cautious line in Broadcom's SEC filing revealing volume floors and escape clauses. Annualized revenue sits at $30B, up roughly 3x in three months.

Anthropic pitches custom silicon as the logical next step for inference cost control and strategic independence; the five-plus-year timeline and absence of any visible chip team expose it as investor theater that cannot ship before 2031.

What Nobody at the Company Can Say

Claude finally became good enough for real enterprise volume, which means inference spend is now the dominant cost line at $30B ARR, and neither Google TPUs nor Nvidia GPUs will ever let Anthropic show gross margins above 62 percent without its own silicon. Every extra nine nines of agent reliability drives exponential token burn, turning the $500B valuation into a leveraged bet on hyperscaler discounts that are already tightening. The Broadcom SEC language is not cautious; it is a polite warning that Google will not subsidize Anthropic's growth forever.

The Engineer Who Quit

The absence of any announced chip architecture hire or tapeout program is the real signal. There is no RTL team, no PD flow, no tapeout budget visible in any public job posting or industry filing as of April 2026. The Reuters leak itself is the departure signal: it exists to calm investors worried about Nvidia pricing power and TSMC packaging scarcity, not to announce an actual hardware program.

Who Pays

Enterprises

Starting Q2 2026

Multi-year Claude Managed Agents subscriptions at 5-10x prior API spend, with no option to source Claude inference elsewhere

Google and Broadcom

Through 2031

Prepaid TPU capacity contracts in low single-digit billions annually through 2031, with take-or-pay volume commitments

Secondary investors

Throughout 2026

Share purchases at $500B-plus valuation with Polymarket pricing 93.5 percent odds of a $500B-plus close in 2026

Dead Pool Watch

The in-house silicon project joins the dead pool by Q4 2026 if zero ASIC job postings appear. Watch for Amazon to acquire once Jassy's chip unit needs a flagship $30B ARR customer to validate an external chip sales narrative. Google already has the TPU hook through 2031 and can wait for a cheaper distressed price.

In 6 Months

By October 2026 Anthropic will have hired at most fifteen chip architects for research purposes and will quietly extend Nvidia and Amazon contracts for another 24 months while de-emphasizing all silicon language

Signal Absence of any tapeout budget line item in board decks and continued heavy promotion of Claude Managed Agents revenue metrics as the primary growth narrative

Gross margins compress below 55 percent as the 3.5 GW TPU deal ramps and Nvidia maintains pricing amid TSMC packaging lockout

Signal Any earnings-style disclosure showing inference cost per token rising quarter-over-quarter despite revenue growth confirms the margin squeeze

What Would Change This

A surprise joint-development agreement with Broadcom to fork the existing TPU architecture for Claude-specific operators, or the immediate hiring of a 200-person Nvidia architecture team backed by $2B in restricted stock grants, would materially alter the trajectory. Absent either move, Anthropic remains a high-end SaaS wrapper on top of rented hyperscaler silicon with investor capital substituting for sustainable gross margins.