Anthropic confirmed on August 5, 2026 that it is building an internal custom silicon team to design chips specifically for its Claude models, according to Inside (inside.com.tw) [E1]. Reuters was first to report the plan, with Anthropic confirming it publicly afterward [E2]. Roughly two months after the initial report, Anthropic separately confirmed to Business Insider that it was forming an internal chip-design team and reiterated it would pursue a "multi-chip approach" going forward, per TechNews [E21].
Anthropic has stressed that custom silicon is additive, not a replacement. The company said it will continue using processors from Amazon Web Services (AWS), Google (谷歌), NVIDIA (輝達), and AMD (超微), according to Inside [E3]. That framing is echoed in Anthropic's statement to Business Insider that it is sticking with a "multi-chip approach" even as the in-house design team takes shape [E21].
Inside reports that Anthropic's official careers page lists a "Silicon Engineer" role based in San Francisco, New York, and Seattle, with an annual salary range of $320,000 to $485,000 [E4]. That single posting is recruiting simultaneously across eight technical areas: front-end design, pre-silicon verification, physical design, design-for-testability (DFT), analog and mixed-signal, process and fab, computer-aided design (CAD) and design infrastructure, and packaging with signal/power integrity [E5].
TechNews reports a pay gap between the engineers who will actually design Anthropic's chips and the researchers building AI systems to assist with that design. A research engineer working on getting AI models to participate in chip design is listed at $500,000 to $850,000 a year (about NT$15 million to NT$25.5 million), compared with $320,000 to $485,000 (about NT$9.6 million to NT$14.55 million) for the engineer developing Anthropic's first ASIC — a gap of up to $365,000 at the top end [E18].
| Role | Salary range (USD) | Salary range (NTD, approx.) |
|---|---|---|
| Silicon/chip engineer (first ASIC) | $320,000–$485,000 | NT$9.6M–NT$14.55M |
| Research engineer, AI chip design | $500,000–$850,000 | NT$15M–NT$25.5M |
The gap was also flagged on social media: the account Papa Johns (@SVTrivo) pointed out that Anthropic's "Research Engineer, Chip Design RL" posting — which builds reinforcement-learning environments so Claude can learn tasks like RTL generation, verification, and physical-design optimization — is listed at $500,000–$850,000 [E19].
According to Inside, Clive Chan — described as OpenAI's custom-chip team's second hardware hire — joined Anthropic in early June 2026 with the title Member of Technical Staff [E6].
Anthropic's own disclosures show fast revenue growth in the run-up to the chip announcement. In an April 6, 2026 newsroom post, Anthropic said annualized run-rate revenue had surpassed $30 billion, up from about $9 billion at the end of 2025 [E7]. By May 28, 2026, when Anthropic announced a $65 billion Series H round at a post-money valuation of $965 billion, run-rate revenue had crossed $47 billion — roughly a 4x increase in about five months. Micron (美光), Samsung Electronics (三星電子), and SK Hynix (SK海力士) all participated in that round as strategic infrastructure partners [E8].
| Date | Metric | Value |
|---|---|---|
| End of 2025 | Run-rate revenue | ~$9B |
| Apr 6, 2026 | Run-rate revenue | >$30B |
| May 28, 2026 | Run-rate revenue | >$47B |
| May 28, 2026 | Series H round | $65B |
| May 28, 2026 | Post-money valuation | $965B |
Supply-side pressure shows up directly in pricing. Qualcomm (高通) notified customers on July 24, 2026 that it would raise chip prices by a double-digit percentage starting September 1, 2026, citing tight semiconductor and memory supply, per Inside [E9]. Further upstream, NVIDIA (輝達), AMD (超微), and hyperscale cloud providers have already booked ABF substrate capacity through 2028 and are asking suppliers to plan capacity expansions for 2029 and 2030 [E14].
OpenAI and Broadcom unveiled their first custom inference chip, Jalapeño, on June 24, 2026, taking just nine months from design to production and manufactured by TSMC (台積電), according to Inside [E12]. Broadcom CEO Hock Tan said early test results showed cost savings of up to about 50% compared with common AI GPU setups [E12].
That cost gap lines up with a broader shift TrendForce is tracking: cloud providers' self-designed ASIC shipments are forecast to grow 44.6% in 2026, compared with 16.1% growth for GPUs — the first time in the AI era that ASIC growth has outpaced GPU growth. ASIC-based AI servers are projected to reach 27.8% of total AI server shipments in 2026, rising to 40% by 2030 [E13].
| Metric | ASIC | GPU |
|---|---|---|
| 2026 shipment growth (TrendForce forecast) | 44.6% | 16.1% |
| Share of AI server shipments, 2026 | 27.8% | — |
| Share of AI server shipments, 2030 (forecast) | 40% | — |
According to TechNews, Anthropic is in talks with Samsung Electronics (三星電子) to manufacture its custom chip using a 2-nanometer process and advanced packaging facilities, though the plan remains at an early stage — even the processor's intended function, target performance, and server integration are still unsettled [E15]. That follows an earlier signal: in May 2026, Anthropic referenced "logic chips" while discussing its ongoing relationship with Samsung, which fueled outside speculation about joint ASIC development [E20].
On the memory side, SK Group Chairman Chey Tae-won (崔泰源) confirmed in July 2026, appearing alongside Anthropic CEO Dario Amodei at a San Francisco AI event, that Anthropic had approached SK Hynix (SK海力士) with a request for component supply [E16].
Beyond custom silicon, Anthropic has been locking in capacity from existing partners. On April 6, 2026, it signed a deal with Google (谷歌) and Broadcom for 5GW of next-generation Tensor Processing Unit (TPU) capacity, expected to come online starting in 2027, with the bulk of it located in the United States [E10]. On August 4, 2026, Anthropic signed a separate six-year, $10 billion contract with AI cloud startup Volta — a company just over six months old — under which Bitdeer will build a 133MW hydro-powered facility in Tydal, Norway, using NVIDIA's (輝達) latest Vera Rubin architecture [E11].
Per Inside, Volta's Norway facility is being built in two phases: phase one is targeted for December 31, 2026, and phase two for March 31, 2027 [E17].
TechNews points to an early signal of where AI-assisted chip design could go: Moonshot's Kimi K3 model has reportedly completed an autonomous chip design in 48 hours, using open-source EDA tools and a self-built GPU compiler to produce a design with roughly 4.0 mm² of area on the Nangate 45nm standard-cell library, achieving a simulated decoding throughput of more than 8,700 tokens per second [E22].
The numbers Anthropic itself has disclosed suggest the custom-silicon push is arriving alongside, not instead of, heavy external chip spending. Run-rate revenue roughly quadrupled from about $9 billion to more than $47 billion in about five months [E7][E8], the same window in which Qualcomm raised chip prices by a double-digit percentage and ABF substrate capacity got booked out through 2028 [E9][E14] — a backdrop that helps explain why Anthropic is simultaneously building an in-house team, signing a 5GW TPU deal with Google and Broadcom, and committing $10 billion to Volta's Norway facility [E10][E11]. Anthropic's own pay scales also reveal where it currently places the highest value: the research role focused on getting AI to help design chips is priced at up to $850,000, roughly $365,000 above the top of the range for the engineer building the company's actual first ASIC [E18][E19] — a gap that sits alongside industry data showing ASIC shipments are now forecast to outgrow GPU shipments for the first time [E13], and alongside OpenAI's own Broadcom-built Jalapeño chip claiming close to 50% cost savings over GPUs [E12].
$320,000 to $485,000 a year, for positions based in San Francisco, New York, or Seattle, according to Inside.
Eight: front-end design, pre-silicon verification, physical design, DFT, analog/mixed-signal, process and fab, CAD and design infrastructure, and packaging with signal/power integrity.
August 5, 2026, after Reuters first reported the plan.
Up to about 50%, according to Broadcom CEO Hock Tan, based on early test results.
$965 billion post-money, following a $65 billion Series H round announced May 28, 2026.
According to TechNews and CNA, Winbond Electronics (華邦電) GM Chen Pei-ming warned on August 6, 2026 that 2027 memory shortages will exceed 2026's. CMS revenue rose 78% and Flash revenue rose 65% quarter-over-quarter, backed by NT$39.5 billion in 2026 capex, up over 70-fold year-on-year.
According to Technews, MoneyDJ, and CTEE reports dated August 6, 2026, TSMC (台積電) is holding roughly $1 billion (about NT$32 billion) worth of Apple (蘋果) A20 Pro processors that cannot complete back-end packaging because of a DRAM shortage. TSMC's CFO had already flagged rising inventory days on July 16, and Apple sources memory mainly from Micron, plus SK Hynix and Samsung.
According to Technews, also reported by CNA and ETtoday, the US FCC is drafting a measure to ban imports of new Chinese-made optical transceivers, with officials hoping to finalize it this year, though it could still be revised or shelved. The news drove same-day share drops of 7.27% for Innolight (中際旭創) and 5.29% for Eoptolink (新易盛), while Beijing's foreign ministry called the move protectionist.