Newsletter · · Ashutosh Agarwal

Anthropic Goes Custom Silicon as AWS Trainium Sells Out Through 2027 - Custom Silicon vs Nvidia - Week of August 10, 2026

For the week of August 3 to 10, 2026, the last big AI lab without its own chip program started one, the head of the world's largest cloud put hard numbers on how sold out his in-house accelerators are, and the custom-silicon story finally moved from the spenders back to the chips themselves.

Custom Silicon vs Nvidia

Week of August 10, 2026: Anthropic Goes Custom Silicon as AWS Trainium Sells Out Through 2027

A note before we start. Last week was the great earnings sort, the market decided to stop treating "AI spending" as one trade and started grading each hyperscaler on whether the money is visibly working. This week the story moved from the spenders to the chips themselves, and it was a genuinely busy one. After a couple of quiet weeks on actual silicon news, we finally got new developments to trade on: the last big AI lab without its own chip program started one, the boss of the world's largest cloud put hard numbers on how sold-out his in-house chips are, Nvidia's next flagship quietly lost some of its memory, and AMD went from "second place" talking point to a company that just signed up every major AI lab at once.

As always, I've tagged each item so you know exactly what you're holding. **** means a company's own executive said it on the record. [REPORTER] means a journalist broke it as reported news. **** means an analyst, strategist, VC, or host is giving an opinion. Design-win and roadmap leaks rank above earnings color, which ranks above hot takes.

One honest sourcing note up front: the two best sources this week were both on broadcast interviews, AWS chief Matt Garman on Bloomberg, and AMD CEO Lisa Su on CNBC. Everything genuinely new about Broadcom, Marvell, Astera Labs, Arm, and Alchip design wins is still missing from the podcast universe, that quiet zone is now in its fifth week, and I'll keep flagging it rather than papering over it.

TL;DR

  • Anthropic is building its own AI chips, the last big lab to go custom. A daily AI news podcast reported that Anthropic is standing up a custom-silicon design team to build its own hardware for Claude and is "scouting Samsung as a manufacturing partner," which "ends Anthropik's status as basically the last major AI lab without a public silicon program." The host's framing of why everyone is doing this: "the big reason is everyone wants to get off of NVIDIA… nobody wants to feel bottlenecked by one player" (AI Chat, 2026-08-05). [REPORTER/PUNDIT]

  • AWS says its own chips are a $25 billion business and sold out through 2027. AWS CEO Matt Garman, on Bloomberg, confirmed the in-house-chip capacity business (Trainium plus Graviton) is running at a "$25 billion" revenue run rate, that "we're largely sold out through the end of next year for Tranium capacity," and that customers "can oftentimes save 20%, 30% off of their inference costs when they run that on Tranium." He'd "definitely consider" selling the chips outright later (Bloomberg Talks, 2026-08-03). ****

  • AMD landed all five frontier labs on one platform, and Lisa Su guided data-center growth "well over 100%" into 2027. AMD unveiled its Helios rack system as a direct Nvidia competitor; Microsoft, OpenAI, Meta, Oracle and Anthropic have all committed to gigawatt-scale deployments, and Anthropic signed for up to two gigawatts on Helios (AI Chat, 2026-08-05). On CNBC, Su said the data-center business is "growing well over 100%" into 2027 and sized the total computing market at "over $2 trillion" by 2030, with Cramer noting AMD is "up 142 percent this year. NVIDIA is up 13 percent" (Squawk on the Street, 2026-08-05). ****

  • Nvidia is quietly putting less memory in its next flagship. The Information reported Nvidia is testing versions of Rubin Ultra with 256GB and even 192GB of memory, down from the "one terabyte of memory per chip" Jensen Huang promised in early 2025, because of an acute high-bandwidth-memory shortage. The twist: less memory per chip means customers "would need more chips," which "honestly is kind of good for NVIDIA" (The Information's TITV, 2026-08-06). [REPORTER]

  • AMD bought a Toronto inference-chip startup that etches models into silicon. Talos, founded 2023, doesn't store model weights in memory, it "etches them directly into the silicon." Its first test chip served Llama 3.1 8B at "16,960 tokens per second," which when announced was "48 times faster than NVIDIA's GPUs and 8.5 times faster than Cerebrus" (Tech Brew Ride Home, 2026-08-07). [REPORTER]

  • The financing plumbing keeps getting scarier, and Nvidia is now the story. A strategist flagged that unsold Blackwell chips are "sitting in warehouses," that the capex-to-sales ratio "just turned down at the beginning of June," and that Nvidia has "invested tens of billions of dollars… in their customers, basically… backstopping their customers." Separately, Nvidia's own credit-default swaps "spiked to a record this month" (RiskReversal, 2026-08-07; Unf*cking The Republic, 2026-08-07). ****

What's new

Ranked by what actually moves a book: the structural chip news first, then the operator color, then the roadmap leaks, then the financing signals.

1. Anthropic is building its own AI chips, closing the loop on the whole ASIC thesis. [REPORTER/PUNDIT] This is the most thesis-defining item of the week. On the AI Chat podcast, host Jaeden Schafer walked through reporting that "Anthropik is building their own custom silicone team to design their own AI chips," creating "custom hardware for Claude," and is "scouting Samsung as a manufacturing partner, which is ending Anthropik's status as basically the last major AI lab without a public silicone program" (AI Chat, 2026-08-05).

Why it matters: it completes the set. As the host put it, "It's the same thing that OpenAI, Google, and Meta, they're all doing it right now. And the big reason is everyone wants to get off of NVIDIA. These companies are incredibly dependent on NVIDIA. NVIDIA basically has quotas and they dole out their GPUs… and nobody wants to feel bottlenecked by one player." He ran the roster: "OpenAI shipped their Broadcom-designed chip, which is called Jalapeno. It's an inference chip, and that was in June of this year. Google now runs their own TPUs. Meta is building MTIA accelerators. And Anthropik basically was the only one that wasn't working on this." Now it is.

The reality check, which he was refreshingly honest about: custom silicon "typically takes between 18 to 24 months to design, and it's going to cost hundreds of millions of dollars per chip version." Anthropic is "betting on a two-year runway before any of that investment is going to get paid back." So this is a margin play, not a shortcut, "Anthropik can tune memory and math formats to Claude's exact workloads in a way that no off-the-shelf GPU can really match." Two caveats for the book: first, this is a media report of a scoop, not something Anthropic said on the record, so treat the program as the news and the timeline as soft. Second, and this is a read-through worth chewing on, the reported foundry partner is Samsung, not TSMC, and there's no mention of Broadcom or Marvell as the design-services partner the way OpenAI used Broadcom for Jalapeño. If Anthropic is building an in-house design team rather than hiring a merchant ASIC house, that's a subtle negative for the "arms-dealer designers win either way" story. Unclear for now; worth watching who they actually sign.

2. AWS put real, specific numbers on its in-house silicon, and they're better than last week's fuzzy "triple digits." This is the highest-quality source of the week, because it's an actual executive on the record. AWS CEO Matt Garman, interviewed on Bloomberg, was asked point-blank what the "$25 billion" chip run rate means. His answer: it's "renting out capacity based on those chips… we run all of our own chips inside of the AWS cloud. And that includes both Tranium chips as well as Graviton chips" (Bloomberg Talks, 2026-08-03).

The lines that move numbers:

  • Sold out: "As we've mentioned a couple of times now, we're largely sold out through the end of next year for Tranium capacity. And it's really, really popular." Being sold out through the end of 2027 is a demand signal you can build an estimate around.

  • The cost advantage, quantified: asked to make it tangible, Garman said customers "can oftentimes save 20%, 30% off of their inference costs when they run that on Tranium." That is the TCO gap in the company's own words, the whole reason a hyperscaler builds its own chip.

  • Trainium 3 is live and popular: "our Tranium 3 chips are really, really popular with customers right now."

  • Why they still buy Nvidia: "we're one of NVIDIA's absolute largest customers in the world… AWS is the most scalable and secure and stable place to run NVIDIA servers anywhere." The pitch is choice, not replacement, Trainium for cost-sensitive, massive-scale inference where AWS controls "that whole stack," Nvidia "for when you need some of that processing for GPUs as well."

  • The future optionality: would AWS ever sell the chips outright rather than rent capacity? "We've mentioned that we might consider that in the future… it's something that we would definitely consider." Not today, "we have so much demand inside of the AWS cloud", but the door is open. If AWS ever merchant-sells Trainium, that's a direct shot at Nvidia's addressable market.

3. AMD stopped being a "second-place" talking point and became a credible Nvidia alternative. Two separate podcasts stacked into one story. First, the product: AMD unveiled Helios, a rack-scale system pitched as a direct competitor to Nvidia's largest AI training clusters. Per AI Chat, "Microsoft, OpenAI, Meta, Oracle, and Anthropik have all committed to deploying it at a gigawatt scale," shipments start "later this year," and Anthropic specifically signed "to deploy up to two gigawatts of GPUs on Helios, which is the exact same scale of NVIDIA's largest AI training cluster." AMD also introduced Venice X, a data-center CPU to pair with Helios, launching 2027. The kicker: "This is the first time that AMD has landed all five major frontier AI labs on a single rack platform" (AI Chat, 2026-08-05).

Then the operator, live on CNBC. Lisa Su said the data-center business is "accelerating," that AMD "just added Anthropic to the mix" alongside OpenAI and Meta as "large strategic partners," and, the number analysts apparently under-modeled, "we do see our data center business growing well over 100% as we go into 2027… the notion of over 100% should [be] considered be well over 100%." She sized "the market for overall computing and adaptive, high performance and adaptive computing going up to over $2 trillion as we go through 2030." She grounded the confidence in visibility: "these data center buildouts are really long-term buildouts. So we have to plan with our customers… 12, 24, 36 months in advance." And she suggested 2027 supply will be easier: "I think as a semiconductor industry, we are really good at planning things… what we're seeing in 2027 is our customers are giving us much more opportunity to plan" (Squawk on the Street, 2026-08-05). Cramer's framing of the scoreboard: "You're up 142 percent this year. NVIDIA is up 13 percent." Note this is merchant-GPU competition, not custom ASIC, but it's the same force (buyers wanting an alternative to Nvidia) and it's clearly working for AMD.

4. Nvidia is quietly downgrading Rubin Ultra's memory, a rare crack in the roadmap. [REPORTER] The Information's Nvidia reporter Phoebe Liu broke that Nvidia "is testing versions of its next-generation GPU, so Rubin Ultra, with less memory than it initially planned." The context: "in early 2025, Jensen Huang said that Rubin Ultra would have one terabyte of memory per chip… a lot more than any chip anyone has ever produced." Now, because of an acute high-bandwidth-memory shortage, partly self-inflicted, since "NVIDIA is buying so much advanced high-bandwidth memory that it's just hard for the suppliers to make enough of it", Nvidia is "testing versions with 256 gigabytes of memory and even 192 with some customers," though "nothing has been finalized yet" (The Information's TITV, 2026-08-06).

Here's the counterintuitive part, and why it's not straightforwardly bearish for Nvidia: with less memory per chip, "to run a model with the same level of performance… you would need more chips. Which honestly is kind of good for NVIDIA because they could also in turn sell more GPUs." Customers Liu spoke to "don't think it will affect demand" and expect the lower-memory chips to be "cheaper." The pressure valve is a "$500 billion partnership with the SK Group" (parent of memory maker SK Hynix) to co-develop next-gen HBM. Rubin Ultra production is "set to start late next year." The read for our beat: the binding constraint on the entire AI buildout, GPUs and ASICs alike, is increasingly memory, not logic. And Liu noted the big clouds are "getting more and more worried about how much they're spending on NVIDIA GPUs", which is, of course, the exact anxiety that funds the custom-silicon programs above.

5. AMD bought an inference-chip startup that bakes the model into the silicon. [REPORTER] Tech Brew detailed AMD's acquisition of Talos, a Toronto startup founded in 2023, "an actual acquisition rather than an acqui-hire." Talos's approach is radically different from GPUs or from Groq's and Cerebras's designs: "The startup's chips don't rely on HBM to store the model weights, but rather etch them directly into the silicon. In a sense, Talos' chips are really model-specific integrated circuits, or MSICs." The proof point: its first test chip, the HC1 on "TSMC's 6nm process," served "Meta's Llama 3.18b at a blistering 16,960 tokens per second," which "was 48 times faster than NVIDIA's GPUs and 8.5 times faster than Cerebrus' accelerators" when announced in February (Tech Brew Ride Home, 2026-08-07).

The catch is the whole point of the debate: "Once the chips are deployed, you're stuck with that model. Any change bigger than something like a LoRa adapter is going to require a re-spin of the chips." With new models "rolling out on a nearly monthly basis," that's a real constraint, though Talos claims a re-spin only requires changing "two layers of metal," not starting over. The second-gen HC2, "due out this summer," targets 20 billion parameters per chip, meaning "you'd need just 50 accelerators to support a trillion-parameter model", versus Nvidia's newly unveiled LPX systems, which "would need a few dozen GPUs and at least 2,000 Grok LPUs to serve the same model." Two things for the book: (1) this is the same extreme-specialization bet as Etched (last week's $300M raise), now with a big-cap acquirer behind it, and (2) it says AMD, like Nvidia with Groq, is buying its way into purpose-built inference rather than betting the GPU can do everything.

6. Google's next custom chip, "Frozen V2," got specs, and Google's AI org got a shakeup. [REPORTER] On the chip side: Google "is building a new AI server chip called Frozen V2. It's going to run the Gemini models, and apparently it's going to be six to ten times more efficient" versus its current chips, arriving "in 2028," with efficiency measured "by tokens generated per watt of power." The report noted Google "has been designing [custom AI chips] internally since 2015" (AI Chat, 2026-08-04). This is the same "Frozen" branch that leaked two weeks ago, now with a name, a year, and an efficiency claim attached, hold the 6–10x loosely, as it's a 2028 target.

On the org side, and this is the read-through: The Information reported Demis Hassabis is stepping aside to become chair of DeepMind and chief scientist at Alphabet, with CTO Koray Kavukçuoğlu taking over the lab. Separately, longtime AI leaders including Jeff Dean, the man who "created the TPU", are leaving to found a startup called Discovery Loop (The Information's TITV, 2026-08-06). The Information's editors downplayed it ("I wouldn't be that concerned about his departure"), and a backing VC pointed out Google still has "their own mature accelerator chip platform in the TPUs." But when the person who invented your custom accelerator walks out the door the same week you're leaning on that accelerator as a competitive weapon, it belongs in the risk column.

7. The financing signals turned sharper, and pointed at Nvidia itself. This is opinion, not a chip fact, but it's the tide under every name here, and this week it got specific about Nvidia. Morgan Stanley's Mike Wilson, on RiskReversal, made three points worth writing down. First, inventory: "there's tons of supposedly Blackwells just… sitting in warehouses that have not been deployed yet," raising the question of whether Nvidia is "stuffing the channel." Second, the rollover: "we've shown this capex to sales ratio factor. It was going straight up over the last year. It just turned down at the beginning of June." Third, the circularity: "NVIDIA has invested tens of billions of dollars over the last few years in their customers, basically… They're backstopping their customers," and he warned the setup "has the potential to be a combination of the dot-com bust and the financial crisis wrapped up in one" (RiskReversal, 2026-08-07). His timing caveat, to be fair: "I just don't think that we're at that stage yet."

The bond-market echo showed up everywhere. Unfcking The Republic* noted "NVIDIA's own credit default swaps… spiked to a record this month," that hyperscaler bond spreads "have widened across nearly every maturity," and that the circular-financing pattern is now so pronounced "even the IMF and the Bank for International Settlements have flagged it as a systemic risk" (Unf*cking The Republic, 2026-08-07). Ed Yardeni, on The Compound and Friends, was more constructive but didn't dismiss it: hyperscalers will spend "an estimated trillion dollars next year," Google "just raised $25 billion worth of debt" against "$100-something billion worth of demand," but hyperscaler bond spreads over Treasuries are "rising like one of their share prices," and "credit default swaps on Oracle… are at like a multi-year high" (The Compound and Friends, 2026-08-07).

8. The spending itself keeps ratcheting up, now $1 trillion for 2027. For the running scoreboard: Carson Group's Sonu Varghese laid out that the five big hyperscalers (Microsoft, Google, Amazon, Meta, Oracle) have seen their 2027 capex estimate climb "from $890 billion to $1 trillion… close to 3% of GDP," with 2026 individual figures at Microsoft "$160 billion… a 93% increase," Alphabet "$200 billion" (a "120% increase"), Meta "$137 billion" (~96%), and Oracle "$77 billion" (116%) (Facts vs Feelings, 2026-08-05). His thesis is the mirror image of the bear case: "all the spending is going to accrue to the revenue and profits of other companies… for now, it's going to be the chip companies." Same dollars, opposite conclusion, which is exactly the debate.

The debate

For the custom-silicon side (ASICs take share and cap Nvidia's TAM and margins). The bull case had its strongest week in a while, and for once it wasn't just vibes. First, the field is now complete: with Anthropic building its own chips, every major AI lab and hyperscaler, OpenAI (Broadcom's Jalapeño), Google (TPU/Frozen V2), Amazon (Trainium), Meta (MTIA), and now Anthropic, has an in-house program, all for the same stated reason, that "everyone wants to get off of NVIDIA." Second, we got an operator quantifying the payoff for the first time in a while: AWS's Garman says Trainium saves customers "20%, 30%" on inference and is "sold out through the end of next year", that's not a science project, that's a capacity-constrained business. Third, the merchant alternative is landing punches too: AMD signed all five frontier labs onto Helios and guided data-center growth "well over 100%," while its stock is up 142% against Nvidia's 13%. Fourth, the constraint that hurts Nvidia most, memory, is so tight that Nvidia is shrinking the memory on its own flagship, and the clouds are "getting more and more worried about how much they're spending on NVIDIA GPUs."

For the merchant side (Nvidia keeps the pie). The steel-man is real. Start with the memory story: even a downgraded Rubin Ultra is arguably good for Nvidia, because fewer gigabytes per chip means customers "would need more chips." Nvidia keeps buying its way into the inference threat (Groq), just as AMD just did with Talos, the incumbents are absorbing the insurgents, not being replaced by them. The specialized-chip bet still carries brutal timing and lock-in risk: Talos chips are frozen to a single model in a world where models change monthly, and Anthropic's own silicon is on a two-year payback runway with the chip that invented Google's TPU (Jeff Dean) now leaving the building. And the CUDA/systems moat is intact where it counts, as All-In relayed, Jensen argues closed frontier models remain cheaper all-in once you count "the training costs and a lot of expertise to fine tune and maintain," which is "why they continue to run away with it on the revenue side" (All-In, 2026-08-08).

The wildcard both sides funnel to: memory and money. Two constraints now sit above the logic-chip debate. Memory is the physical bottleneck, Yardeni's point that the real earnings winners are Micron and Western Digital, "selling out of their supply" at "60%, 70%" margins, applies whether the chip on top is a GPU or an ASIC. Financing is the other, with Nvidia's CDS at a record and the capex-to-sales ratio rolling over in June, the dollars underwriting the whole buildout are getting more expensive for everyone. Dearer money favors the cheapest chip that does the job (the ASIC argument); but if it chokes off the buyers, nobody's chip ships.

Where I come out this week: the ASIC thesis got its best confirmation yet, Anthropic completing the field, and AWS finally quantifying the cost win and admitting it's sold out. That's the tell. But the same week gave the bears fresh ammunition on Nvidia's financing rather than its technology, CDS at records, Blackwells in warehouses, which is a different, and in some ways scarier, risk than "someone builds a better chip." I'm holding my constructive lean on the in-house-silicon franchises (Google TPU, Amazon Trainium) and adding AMD to the "credible merchant alternative" bucket, while treating memory (HBM) as the highest-conviction picks-and-shovels exposure and watching Nvidia's credit spreads as the single most important dial on the board.

Stocks in play

  • NVDA (Nvidia). [REPORTER + PUNDIT] Bull: even a lower-memory Rubin Ultra could mean customers "need more chips," so Nvidia "could also in turn sell more GPUs"; the $500B SK Hynix partnership secures memory supply; Jensen's argument that closed frontier models stay cheaper all-in keeps the revenue running (The Information's TITV, 2026-08-06; All-In, 2026-08-08). Bear: Rubin Ultra's memory cut from 1TB to as low as 192–256GB is a rare roadmap concession to a shortage it helped cause; Blackwells reportedly "sitting in warehouses"; capex-to-sales rolled over in June; CDS at a record; "backstopping their customers" to the tune of "tens of billions" (RiskReversal, 2026-08-07). Watch: final Rubin Ultra memory spec and pricing; whether channel inventory is real demand or stuffing; the CDS trend.

  • AMZN (Amazon). Bull: in-house chips (Trainium + Graviton) at a "$25 billion" run rate, Trainium "sold out through the end of next year," 20–30% inference cost savings, Trainium 3 "really, really popular," and the option to merchant-sell chips later (Bloomberg Talks, 2026-08-03). Bear: AWS remains "one of NVIDIA's absolute largest customers," so it's a hedge, not a replacement; capex ~$220B with negative free cash flow (recap from last week). Watch: whether AWS ever sells Trainium outright, that would be a direct hit to Nvidia's TAM.

  • AMD. Bull: first to land "all five major frontier AI labs on a single rack platform" (Helios), Anthropic signed for up to 2GW, data-center growth guided "well over 100%" into 2027, and the Talos acquisition adds a purpose-built inference weapon; stock +142% YTD (AI Chat, 2026-08-05; Squawk on the Street, 2026-08-05). Bear: the stock has run hard, so it's priced for execution; Helios ships "later this year" and the 2GW deployments are 2027, delivery risk is real. Watch: Helios shipment timing and any concrete Trainium/TPU-style deployment numbers.

  • GOOGL (Alphabet). [REPORTER] Bull: "Frozen V2" custom chip for Gemini, "six to ten times more efficient," coming 2028; TPU remains the reference external ASIC; still "the only totally full stack platform" with "their own mature accelerator chip platform" (AI Chat, 2026-08-04; The Information's TITV, 2026-08-06). Bear: the TPU's creator Jeff Dean and other senior AI leaders are leaving; a leadership reshuffle spooked the stock; ~$200B capex (+120%) with the financing questions that follow. Watch: whether the brain drain slows TPU/Frozen execution, and any external-TPU customer disclosure.

  • AVGO (Broadcom). [READ-THROUGH, indirect] Bull: OpenAI's shipped inference chip "Jalapeno" is "Broadcom-designed," the concrete proof that the merchant-ASIC-design model works at frontier scale (AI Chat, 2026-08-05). Bear: no new AVGO-specific design-win commentary this week; and Anthropic's new program is reportedly leaning on Samsung as a foundry with an in-house design team, not visibly on Broadcom. Watch: whether Anthropic or any new lab picks Broadcom for design services; XPU customer-list updates.

  • MRVL (Marvell) / ALAB (Astera Labs). [READ-THROUGH, coverage gap] Bull: structurally, more custom-silicon programs (now including Anthropic) mean more ASIC-design and connectivity attach over time. Bear: zero company-specific commentary on either this week despite dedicated searches, a persistent quiet zone, now five weeks running. Watch: any design-win disclosure tied to Trainium 3, TPU/Frozen, or Maia.

  • MSFT (Microsoft) / META. Bull/Bear: no fresh Maia or MTIA chip detail surfaced this week; the commentary was a rehash of last week's capex-discipline split (Azure strength vs. Meta's free-cash-flow collapse). Notably, one bear argued "70% of Microsoft's AI revenues" are tied to OpenAI (Better Offline, 2026-08-07). Watch: any Maia 2 or MTIA v2 deployment number, still a black box.

  • ARM. [COVERAGE GAP] No Arm-specific podcast commentary surfaced this week, no Neoverse-in-ASICs or v9 royalty color. Flagging honestly rather than fabricating. Watch: royalty-mix disclosure as hyperscaler ASIC volumes ramp.

  • Private/adjacent: Anthropic, Talos, Etched, Groq, Cerebras. [REPORTER/PUNDIT] This is where the week's genuinely new custom-silicon energy sat: Anthropic starting its own program (Samsung foundry), AMD buying Talos (MSICs, 48x-vs-GPU benchmark), following last week's Etched raise and Nvidia's Groq deal. Watch: Talos HC2 (due this summer, 20B params/chip) and whether Anthropic names a design/foundry partner.

Read-throughs

  • Memory (HBM) is the read-through of the week, and the toll booth just got more valuable. Nvidia is cutting Rubin Ultra's memory because it can't get enough HBM, and threw "$500 billion" at SK Hynix to fix it (The Information's TITV, 2026-08-06). Yardeni's framing: the real earnings winners are the memory makers, "Sandisk and Micron, which are now gigantic index weights… selling out of their supply" at "60%, 70%" margins, with Western Digital posting "a $3.2 billion quarterly profit… up from $282 million a year ago" (The Compound and Friends, 2026-08-07). HBM (Hynix / Micron / Samsung) gets paid on every accelerator, GPU or ASIC, and it's the one layer where the shortage is unambiguous.

  • TSMC / advanced packaging, a small but real data point. Talos fabbed its HC1 on "TSMC's 6nm process" (Tech Brew Ride Home, 2026-08-07), a reminder that both merchant and custom silicon converge on the same foundry. No fresh CoWoS/CoPoS capacity commentary surfaced, but it remains the shared choke point behind every design named here.

  • Connectivity (Astera Labs, Marvell, Broadcom switching), quiet again. No NVLink/UALink/Ethernet or Tomahawk/Teralynx commentary this week. The structural case (more custom chips need more fabric) is intact, but there was no new signal to trade.

  • Foundry/design partner choice is the emerging sub-plot. Anthropic reportedly went to Samsung as a foundry with an in-house design team; OpenAI used Broadcom for Jalapeño; Talos used TSMC. Which partner each new program picks is becoming the cleanest way to handicap who captures the ASIC-services dollars, watch the pairings, not just the programs.

  • The financing plumbing, still the fault line under everything. Nvidia CDS at a record, hyperscaler bond spreads widening, capex-to-sales rolling over, Google raising $25B of debt, Oracle CDS at multi-year highs (RiskReversal, 2026-08-07; Unf*cking The Republic, 2026-08-07; The Compound and Friends, 2026-08-07). This is the shared risk to every name in the letter.

What changed vs last week

Last week (issue dated 2026-08-03) was the earnings-week capex sort: Microsoft rewarded for disciplined spending, Meta punished for spending without a payoff story, Amazon given a pass because its cloud and its own chips were visibly working, plus the first real callout of Amazon's "triple digit" in-house-silicon growth. This week, four things concretely changed:

  1. The custom-silicon field filled in, Anthropic went from GPU renter to chip designer. Last week the ASIC proof points were Google (TPU) and Amazon (Trainium). This week the last big holdout, Anthropic, reportedly started its own program with Samsung, so every major lab and hyperscaler now has one (AI Chat, 2026-08-05). That's a structural upgrade to the thesis, not just another data point.

  2. AWS put hard numbers where last week we only had "triple digits." Last week's Amazon signal was a vague "triple digit percentages." This week its CEO quantified it: a "$25 billion" run rate for the in-house-chip capacity business, "sold out through the end of next year," with "20%, 30%" inference savings (Bloomberg Talks, 2026-08-03). Much more useful for a model.

  3. AMD emerged as a credible merchant #2. This didn't feature last week. AMD landing all five frontier labs on Helios, signing Anthropic for 2GW, guiding data-center growth "well over 100%," and buying Talos, is a new front in the "alternative to Nvidia" war, and its stock's 142% YTD run says the market already noticed (Squawk on the Street, 2026-08-05).

  4. The bear case moved from macro rates to Nvidia's own credit. Last week the financing worry was generic (2007-high yields, a soft bond deal). This week it got specific and pointed at Nvidia: CDS "spiked to a record," Blackwells "sitting in warehouses," capex-to-sales "turned down at the beginning of June" (RiskReversal, 2026-08-07; Unf*cking The Republic, 2026-08-07).

What did not change: still a coverage gap on Marvell, Astera Labs, Arm, and Alchip (fifth week running); still no published, production-grade ASIC benchmark beating Blackwell in a real deployment (Talos's numbers are on an old, tiny model and carry a model-lock caveat); still no clean external-shipment figures for Maia or MTIA; and networking/connectivity commentary stayed quiet. The honest summary: a genuinely busy week for the chip story, a new lab program, real operator numbers, a memory-roadmap concession, and an M&A benchmark, with the financing risk sharpening around Nvidia rather than the technology.