Newsletter · · Ashutosh Agarwal

Anthropic in Talks With Samsung to Build Its Own Custom AI Chip - Custom Silicon vs Nvidia - Week of July 6, 2026

Custom Silicon vs Nvidia newsletter for the week of July 6, 2026 (June 29 to July 6 coverage window). Anthropic became the eighth credible name in talks to go custom, with Samsung, while every operator on the tape, from Etched to Groq, conceded Nvidia's edge and the moat debate shifted from CUDA to a 30 million installed GPU base.

Custom Silicon vs Nvidia

Week of July 6, 2026: Anthropic in Talks With Samsung to Build Its Own Custom AI Chip

A note on the tape: Thirty episodes touched our beat in the strict 7-day window (June 29–July 6), no 14-day fallback needed. But be honest about the mix: the operators who showed up this week (Etched's CEO, Groq's founder) run merchant challengers, not hyperscaler ASIC programs. On the in-house-silicon story itself, the fresh news is a leak, and the best analysis is a pundit's. Quiet on the design-win front, loud everywhere else.

TL;DR

  • Anthropic is in talks with Samsung to build its own custom chip, the eighth credible name to go custom, and the first fresh "who's-next" datapoint since OpenAI's Jalapeño. Still very early; nothing finalized.
  • Dylan Patel handed us the number that matters: Google runs three separate TPU architectures, one with Broadcom, one with MediaTek, one it won't name, and "everyone will have their own ASIC program... in the case of Google, hundreds of billions of dollars a year."
  • Nvidia is being treated like a value stock. It raised $25B in bonds it doesn't need, the tape calls it "dead money," and the smart-money reframe is that the moat was never CUDA, it's ~30 million installed GPUs.

What's New

Ranked by what moves a book, fresh buyer leaks first, operator color next, pundit framing last.

1. Anthropic + Samsung: the eighth buyer goes custom. [NEWS] Reported this week on AI Chat (Jul 2): Anthropic is "in talks with Samsung to create a custom AI chip... joining OpenAI, Google, Amazon, all of them that are trying to basically reduce the reliance on NVIDIA." The tell for a book is the sobriety of the caveat: "definitely in the early stages... the purpose, the server role, the performance targets, none of that have actually been finalized." Samsung already fabs Nvidia silicon and is building a joint chip fab with Nvidia in Korea, so this is a foundry/memory partner leaning into design, not a Broadcom-style ASIC house win. Treat it as thesis confirmation (the direction), not a 2026 numbers event.

2. Google's three TPU programs, and "hundreds of billions a year." [PUNDIT] SemiAnalysis's Dylan Patel, on Training Data (Jun 30): "Google actually has three different design programs for TPUs. They're making a TPU with Broadcom. That's a different architecture than the TPU with MediaTek... and the third one is a very different architecture." Then the framing bulls have wanted quantified: "everyone will have their own ASIC program... tens of billions of dollars... In the case of Google, hundreds of billions of dollars a year of their own ASICs." Crucially, he keeps the merchant floor intact, "even [if] the majority of the pie goes to NVIDIA and TPU and Tranium." This is the cleanest AVGO read-through of the week, and a reminder that MediaTek is a live second TPU vendor.

3. Etched unveils, a merchant challenger, not a hyperscaler. [OPERATOR] CEO Gavin Iberti on Bloomberg Tech (Jun 30) pitched two tricks: "low voltage inference," "we run out under half the voltage of typical NVIDIA GPUs," and "cluster scale memory," in a rack "pre-assembled with 32 of our chips." He claims "best in the world performance" and lower cost-per-token, at a reported ~$5B valuation, backed by Jane Street, with "around half of our platform team... from NVIDIA," including the VP who ran Nvidia's DGX/HGX line. On Invest Like the Best (Jun 30) he leaned on velocity: "There is another very famous AI chip company that took 10 months to go from getting their silicon back to having them running inference in a rack... We were able to do it in 40 days." Caveat, loudly: benchmarks are the company's own, and customers are "test[ing]," not deploying at scale.

4. Nvidia as a value stock. [PUNDIT] On The Circuit (Jun 29), the hosts flagged Nvidia raising "$25 billion in investment-grade bonds" it plainly doesn't need, "putting a monster pile of cash in reserve," while the stock "is performing like absolute dog shit" and "even the most ardent NVIDIA stock supporters call it dead money." The more durable insight: "CUDA may not actually be the real moat... the real moat is their installed base... 30 million GPUs out there in the world. And the closest person's less than 10, maybe. And that's a custom ASIC." Dylan Patel agrees the "CUDA moat... is at least partially disentangled because models are just great at coding and all software gets commoditized."

5. The TPU's own creator concedes the merchant point. [OPERATOR] Groq founder Jonathan Ross, who built Google's first TPU, on David Senra (Jul 5): "over time, these GPUs have gotten as good and better than TPUs. As the creator of the TPU, I have to admit, GPUs are now better. This is the benefit of just having an entire industry behind you and ecosystem." His LPU-plus-GPU pairing (LPU applies the weights, GPU does attention) is itself an admission that even a custom-silicon zealot buys Nvidia.

The Debate

For ASICs. An eighth name (Anthropic) is going custom, Google is running three parallel TPU designs, and Patel's "hundreds of billions a year" makes the in-house spend too big to dismiss. Software commoditizes, models write their own kernels now, so the historic CUDA lock is loosening at the inference layer.

For the merchant. Every operator this week undercut the panic: Ross admits GPUs beat TPUs; Patel keeps "the majority of the pie" with Nvidia and TPU/Trainium; the real moat is ~30M installed GPUs, not the software. And the "challengers" with cameras this week (Etched, Groq) are startups with self-reported benchmarks and test-stage customers, not hyperscaler volume.

Where I come out this week: the custom-silicon cast list keeps growing while the scoreboard doesn't move, no MLPerf print, no Maia/MTIA/Trainium volume, and Nvidia's own concession is that it's "dead money," not dead. Stay long the picks-and-shovels (AVGO, and the memory complex) and treat NVDA weakness as multiple compression, not thesis break.

Stocks in Play

  • AVGO. Bull: the only merchant name explicitly named this week as a live Google TPU partner (Training Data); Anthropic's move adds another future XPU/connectivity socket. Bear: MediaTek is the second TPU vendor, Broadcom is not the sole beneficiary. Watch: whether Anthropic's chip routes to Broadcom or stays a Samsung captive design.
  • NVDA. Bull: installed base ~30M GPUs; even custom-chip founders still buy it. Bear: "dead money," $25B raised into a flat stock, sentiment pivoting to Micron (The Circuit). Watch: any MLPerf-style benchmark from an ASIC that actually beats Blackwell.
  • GOOGL. Bull: three TPU architectures and "hundreds of billions a year" of in-house silicon is the strongest ASIC franchise on the board. Bear: even Google keeps some workloads on GPUs. Watch: TPU v7 (Gemini's next model is optimized for it, per Patel).
  • AMZN / MSFT / META. Bull: Anthropic (Trainium's anchor tenant) doubling down on custom via Samsung validates the category. Bear: still zero fresh Maia/MTIA/Trainium volume disclosed. Watch: Trainium 3 and any Rainier metric.
  • Samsung / MU / SK Hynix. Bull: Samsung now sits at the center of both fab and design; memory is the toll booth. Bear: Six Five's Daniel Newman insists memory is "by default a commodity" at the JEDEC pin (Six Five). Watch: custom HBM4, the one place memory escapes commodity gravity.
  • QCOM. Bull: data-center debut got real airtime at its investor day (Six Five). Bear/Watch: still no HBM, still no revenue.
  • MRVL / ALAB / ARM. Bear/Watch: no fresh design-win, connectivity, or royalty color this week, a coverage gap, not a negative.

Read-Throughs

  • HBM, Samsung / SK Hynix / Micron. Samsung and SK Hynix are "nearly 80% of the global market for high-bandwidth memory," now underwriting a ~$590B South Korea chip complex; Jefferies sees memory prices +40–50% in Q3 and +30–40% in Q4, with Ming-Chi Kuo warning "15-20% of 2026 consumer electronics capacity" shifts to data centers in 2027 (Tech Brew Ride Home). Every new ASIC pulls this harder.
  • Custom-ASIC memory hoarding. The Circuit flagged "some smoke" that big custom-ASIC players are "securing that memory in advance... to win some deals," i.e., double-ordering to bundle memory with silicon. Watch for air-pocket risk if it's not all real demand.
  • TSMC / advanced packaging. Etched's Iberti on why the foundry wins: "TSMC customer service is way, way better than I have seen at any other company... It is why they are the number one." Packaging and leading-edge capacity remain the shared chokepoint for GPUs and ASICs alike.
  • Connectivity / ASIC services / Arm. No fresh Marvell, Astera Labs, Alchip, or Arm/Neoverse color this week. Watch attach rates as TPU-v7 and the new Anthropic/OpenAI designs land in '27.

What Changed vs Last Week

Last week (June 29) was a product week, OpenAI's named Jalapeño chip with two executives on camera. This week is a list-lengthening week: Anthropic became the eighth name in talks to go custom (via Samsung), but it's early-stage and unfinalized, so the delta is confirmation of direction, not a new benchmark. Two real shifts: (1) the operators who showed up flipped from hyperscaler insiders (Tan, Brockman, Feldman) to merchant challengers (Etched, Groq), and both, notably, conceded Nvidia's edge; (2) the moat debate moved off CUDA and onto the installed base (~30M GPUs) as the harder-to-topple asset. Nvidia sentiment deteriorated further, a $25B bond raise into a "dead money" tape. Still missing, as ever: a published ASIC benchmark that beats Blackwell, and hard Maia/MTIA/Trainium/Marvell volume.