Newsletter · · Ashutosh Agarwal

Amazon's In-House Chips Ramp as the Market Punishes Capex - Custom Silicon vs Nvidia - Week of August 3, 2026

A synthesis of what investor and deep-tech podcasts said about custom AI silicon for the week ending August 3, 2026, including Amazon disclosing triple-digit growth in its homegrown Trainium and Graviton chips, a market that split Big Tech by capex discipline, and the emerging fight over AI networking standards.

Custom Silicon vs Nvidia

Week of August 3, 2026: Amazon's In-House Chips Ramp as the Market Punishes Capex

A note before we start. Last week Alphabet stole the show: Google Cloud grew 82%, the TPU became an external business for the first time, a next-gen chip ("Frozen") leaked, and the stock got punished for its first-ever negative quarter of free cash flow. This week the other giants reported, Microsoft and Meta on Wednesday, Apple and Amazon on Thursday, and the market did something new. It stopped treating "AI spending" as one trade and started grading each company separately: reward the ones who can show the money is working, punish the ones who can't. Microsoft ripped ~10%. Meta fell ~10%. Amazon rose despite hiking its budget by another $20 billion.

For a custom-silicon newsletter, that split matters more than it looks. The whole bet behind in-house chips is that when hyperscalers are forced to care about return on capital, they get cheaper, and cheaper means building your own silicon instead of renting Nvidia's at fat margins. That pressure arrived in force this week, and the single hardest custom-silicon fact of the quarter came with it: Amazon quietly telling us its homegrown chips are now growing at triple-digit rates. Throughout, it is noted whether an item is a company's own numbers or an executive's words, a builder with skin in the game, or an analyst, VC or host opinion, so you know exactly what you're holding.

TL;DR

  • Amazon's own chips are now a real, fast-growing business. On the Thursday call, CEO Andy Jassy said AWS grew 37%, its fastest in 18 quarters, and that AWS's AI business topped a $25 billion annual run rate, growing "triple digit percentages." Crucially, CNBC's reporter flagged that Amazon's "homegrown semiconductor business… surprised to the upside," also "growing triple digit percentages," and analyst Mark Mahaney named "the Trainium and Graviton businesses" as additive on top of that. This is the closest thing to hard custom-silicon momentum we got all week (Fast Money, 2026-07-30).
  • The market split Big Tech by capex discipline. Microsoft held its budget flat (~$175B for the year), showed Azure accelerating to 43%, and rose ~10%. Meta raised the bottom of its range to $130-145B, watched free cash flow collapse to $784 million (from $8.6 billion a year ago), couldn't explain how the spending pays off, and fell ~10% (The Rundown, 2026-07-30).
  • A new inference-chip startup put a number on the ASIC land-grab. On 20VC, the hosts discussed Etched raising $300 million ("3% dilution," implying roughly a $10B valuation) to build an inference-only chip, and pegged the field at "10 or 11 companies… chipping away at three or 400 billion a year of spend." Their CUDA line is the one to remember: "Open doesn't need CUDA. Open is cheaper and it's lower margins. Open will bypass him" (20VC, 2026-07-30).
  • The networking layer is becoming its own gold rush. A deep-tech interview laid out the case that an open "Ethernet Scale-Up Network" (ESUN) will emerge to connect everyone's custom chips, the way Ethernet once won general computing. The numbers: the four biggest clouds have committed more than $700 billion of capex in 2026, up ~77% year on year, and Dell'Oro sees AI back-end network switching passing $100 billion cumulative by 2030, "possibly the largest new market the chip industry has ever seen" (TechSurge, 2026-07-28).
  • Money got more expensive, right as the buildout needs it most. The Fed held rates in a rare split, with three officials wanting to hike, the 30-year Treasury yield jumped above 5.2%, its highest since 2007, and a recent hyperscaler bond deal was only 1.7x oversubscribed versus 5x earlier in the year. Higher borrowing costs squeeze every buyer of every accelerator.

What's new

Ranked by what actually moves a book: the one real custom-silicon data point first, then the capex prints that bear on the thesis, then the pundit debates.

1. Amazon's homegrown chips are growing triple-digits, and it just got a pass for spending more. This is the most thesis-relevant item of the week, and it came straight from Amazon's Thursday-night call. CEO Andy Jassy said AWS grew 37% year over year, its "fastest growth in 18 quarters" (the bar was 31%), and that AWS's AI business "topped 25 billion in annual revenue in terms of the run rate," growing "triple digit percentages year over year." Then the part this newsletter exists to catch: CNBC's Kate Rooney, reading the release, said Amazon's "homegrown semiconductor business… this one surprised to the upside," also "growing triple digit percentages." Analyst Mark Mahaney, quoted on the same show, listed "the Trainium and Graviton businesses" as real, additive lines on top of the cloud number.

A caveat worth holding onto, because the numbers are a little slippery: the $25 billion run-rate is AWS's AI business broadly (Trainium chips plus AI services), and the "$25 billion" figure got echoed for the chips business specifically in the same breath, so treat "triple-digit growth in in-house silicon" as the durable takeaway rather than a precise dollar figure for Trainium alone. Amazon does not break Trainium out cleanly. But the direction is unambiguous, and it's the kind of momentum the TPU showed last week: a hyperscaler's own chips crossing from science project into a business with real revenue behind it.

Why the market let Amazon get away with it: Jassy raised full-year capex to "approximately $220 billion" (from ~$200B), blaming higher memory costs, and said flatly, "we're enthusiastic about the ROIC equation, even with heavy CapEx over the next few years." Free cash flow was negative $7.6 billion over the trailing year and long-term debt "doubled to $129 billion" (analyst Patrick Moorhead), yet the stock rose ~6%. The difference from Meta: Amazon can point at accelerating cloud, a booming AI run-rate, and its own chips growing triple-digits as the payoff. When you can show the money working, the market forgives the spend. The desk did add the standing asterisk: "Anthropic and OpenAI are AWS's two largest customers," so a chunk of that demand is the same two labs everyone else is also leaning on.

2. Microsoft vs Meta: the market rewrote the rules on capex in a single night. For a custom-silicon investor, the interesting thing here isn't the stock moves, it's why they happened, because the "why" is the whole ASIC thesis in miniature.

Microsoft held the line. Revenue rose 18% to $90 billion, net income jumped 31% to nearly $36 billion, and the star was Azure: 43% growth, its fastest since 2022, crossing $100 billion in annual revenue for the first time, with guidance for roughly 45% next quarter. Copilot paid seats passed 30 million, up from 20 million just three months earlier, a 50% jump. And critically, Microsoft did not raise capex: CFO Amy Hood kept the 2026 number around $175 billion, in line with prior guidance, even as spend hit $41 billion in the quarter (up 70% year over year). Jim Cramer, who had been skeptical, capitulated live: "the Microsoft call was fantastic… Azure was extraordinary… Copilot is 30 million seats. They own a distribution system, which AI can be sold through," with a commercial backlog he pegged at "$670 billion… thanks to OpenAI" (Squawk on the Street, 2026-07-30). Stock up ~10%.

Meta did the opposite of what the market wanted. The core business was fine, revenue up 28% to $60.8 billion, but earnings came in at $6.18 a share versus ~$7.22 expected (dinged by a $2.4 billion legal charge and $1.2 billion of severance), next-quarter guidance was light, and the number that scared everyone was free cash flow: just $784 million, down from $8.6 billion a year ago. Meta raised its capex range to $130-145 billion with no clear story for how it pays off, because, unlike Google, Microsoft, and Amazon, Meta has no cloud business to monetize the buildout. Mark Zuckerberg floated becoming an AI-compute seller but wouldn't commit: "We see a large enterprise opportunity to sell to businesses, including APIs, business agents, potentially selling compute directly," while insisting "there will continue to be a significantly higher margin on selling intelligence rather than selling compute directly" (Fast Money, 2026-07-29). Cramer's verdict was blunt: Reality Labs is still burning, "$4.62 billion" spent in the quarter for "$431 million in revs back… cumulative losses there are $80 billion," and "what free cash flow? There is none." Stock down ~10%.

The read-through for our beat: the market is now explicitly rewarding return on capital, not spending for its own sake. That is precisely the environment in which building your own cheaper silicon beats renting merchant GPUs, because the incentive to cut the Nvidia bill just went up.

3. Etched raised $300 million to build an inference-only chip, and the VCs put the whole ASIC bet in plain English. On 20VC, the hosts walked through the case for narrow, inference-only silicon using the cleanest analogy going: 30 years ago, someone realized graphics was "just one piece of math… and people should build a separate chip to do that. And a company called Nvidia did it." The question now, they argued, is whether inference is the same story: "if all you're doing is… just inference, is there an even more narrowly defined chip that in return for giving up on general purpose calculations can be even better for that? Probably is. And that's what [Etched] is making."

The specifics worth writing down: Etched raised $300 million at "3% dilution" (so a roughly $10 billion valuation), and there are "about 10 or 11 companies doing it, all chipping away… at three or 400 billion a year of spend." One host's framing of the payoff is the venture bet in a sentence: "I don't know whether… etched is worth 10 billion or… 200 billion." They were candid that it's hard, Cerebras took "a long 10 years," and that the timing risk is real: "it'll be hard if the timing of tape out happens in a capex decline." The other genuinely new thread was Jensen Huang's pro-open-weights letter (his first tweet in ages), which the hosts read as Nvidia hedging: "Open doesn't need CUDA. Open is cheaper and it's lower margins. Open will bypass him, but he's got it." That's a live crack in the CUDA-moat argument: if open-weight models keep taking share (they noted "almost half of open router's traffic is to open source, open weight models"), Nvidia's software lock-in matters a little less each quarter.

4. The networking layer is turning into its own market, and it's the read-through nobody's pricing. The best deep-dive of the week came from TechSurge, interviewing the founder of Upscale, an AI-networking startup that has raised close to $500 million in under nine months. His pitch: AI compute is going heterogeneous, hyperscalers building their own XPUs alongside Nvidia, and just as Ethernet beat Token Ring, ATM, and fiber channel to become the one standard for general computing, an open standard will win AI networking. The emerging candidate is called ESUN, "Ethernet Scale-Up Network." Today Nvidia's proprietary NVLink dominates, its NVL72 rack connects "72 chips… through NV switch," and Nvidia "own[s] 95% of the market segment," but "the rest of the world doesn't have that option," so a common fabric benefits everyone building custom chips.

Upscale is building its own scale-up switch chip, codenamed "Skyhammer," plus a full networking software stack, and the founder said it is "working very closely with a few hyperscalers on the scale-up switching… with their AI XPUs," aiming to deploy "sometime in the next year." He also validated the inference-ASIC wave from the other side: "NVIDIA's acquisition of Grok really validated" purpose-built inference chips (this is Groq, the inference-chip startup Nvidia struck a licensing deal with, not to be confused with xAI's chatbot).

The market-size numbers are what make this a read-through, not a curiosity. The host laid them out: the four biggest clouds "have committed to more than $700 billion of capital spending in 2026, up roughly 77% in a single year. Most of it buys accelerators. But every accelerator is only as good as the network that feeds it." Dell'Oro expects AI back-end network switching to "blow past $100 billion in cumulative sales by 2030" and calls scale-up networking "possibly the largest new market the chip industry has ever seen." Standards are moving fast: UALink published its first spec in April, Ultra Ethernet followed in June, and in October a 12-company group, AMD, Broadcom, Cisco, Meta, Microsoft, and NVIDIA among them, launched the ESUN effort, with 175 companies joining since. He also flagged the milestones of the week: Nvidia crossed $5 trillion in market cap, and Micron crossed $1 trillion, the market pricing in compute, then memory, as the constraints, with networking arguably next.

5. The macro turned against the whole buildout at once. The Fed held rates at 3.5-3.75% but in a rare split, with three officials wanting to hike because inflation is still above target. New chair Kevin Warsh's press conference was read as muddled, and the bond market recoiled: the 30-year Treasury yield jumped above 5.2%, the highest since 2007, and the 10-year hit 4.7%. Why a chip newsletter cares: this buildout runs on borrowed money. On Fast Money, the desk noted a recent hyperscaler debt offering was "only 1.7 times oversubscribed. That's like nothing in the bond world. There were 5 times earlier in the year," and concluded, "this isn't a good environment for the hyperscalers at the moment." Dearer money is exactly what forces the cost discipline that favors in-house silicon, and also what could stall the whole thing if the buyers can't fund their commitments.

6. The compute-demand bulls fired back. For balance, the loudest counter to the bears came on The Compound and Friends, where Alex Kantrowitz and the hosts argued the cloud re-acceleration is real and broad: Google Cloud +82%, Azure +43% (now $30B a quarter, "insane" on a base three times its 2022 size), AWS +36.7% ("fastest in 18 quarters"). Kantrowitz summarized Amazon's edge as, "do very little, build a lot of compute profit… we have space for it," well positioned if the premium per model compresses and customers route across "five or six" models. They also chewed on Dwarkesh Patel's provocative essay arguing compute could get 10x more expensive, noting Anthropic's revenue "has 10x[ed] year over year" and "likely ends the year with 100 to 150 billion dollars of revenue," though Kantrowitz thinks model convergence ultimately pushes margins down, not up (The Compound and Friends, 2026-07-31). Useful context, but it's opinion, not a new fact about anyone's chip.

The debate

For the custom-silicon side (they take share and cap Nvidia's margins). The bull case got a subtle but real boost this week, and it's structural rather than a single leak. First, the market has flipped from applauding capex to demanding returns, the clearest possible signal that hyperscalers now have to squeeze cost out of their compute, and the fattest cost to squeeze is a merchant GPU sold at premium margins. Second, Amazon gave us the quarter's one hard data point: its own chips (Trainium, Graviton) growing "triple digit percentages," echoing the TPU's crossover last week. Third, the ecosystem around custom silicon is filling in, an inference-chip startup (Etched) just raised $300M into a field of 10-plus competitors attacking a $300-400B market, and the networking layer that stitches custom chips together is being called "possibly the largest new market the chip industry has ever seen." And the CUDA moat took a small ding: if open-weight models keep grabbing share, Nvidia's software lock-in matters a little less, because "open doesn't need CUDA."

For the merchant side (Nvidia keeps the pie). The counter is that this was a week of talk, not silicon. No custom chip shipped, benchmarked, or won a new socket. Nvidia crossed $5 trillion in market cap while the ASIC field debated valuations for startups that haven't taped out. The inference-chip bulls admit it themselves: it took Cerebras "a long 10 years," and a tape-out landing during a capex slowdown could be fatal. Meanwhile the same cloud acceleration that helps the ASIC story, Azure +43%, AWS +37%, is also raw demand that Nvidia captures today, and Nvidia keeps buying its way into the inference threat (the Groq licensing deal). And the standards war cuts both ways: NVLink still owns ~95% of scale-up networking, and an open ESUN standard is years from displacing it.

The wildcard both sides funnel to this week: financing. Last week the swing question was Google's roadmap. This week it's the money. Yields at 2007 highs, a bond deal barely oversubscribed, Meta's free cash flow down to a rounding error, Amazon's negative $7.6B, the dollars underwriting the buildout are getting thinner and dearer. Cheaper money favored spending on whatever chip you wanted; dearer money favors the cheapest chip that does the job, which is the custom-silicon argument. But push it too far and the buyers simply can't fund the commitments, and then it doesn't matter whose chip is more efficient.

Where this lands for the week: the scoreboard didn't move much on the silicon itself, but the rules of the game shifted in the ASIC thesis's favor. The market now punishes undisciplined capex (ask Meta) and rewards proof of monetization (ask Microsoft), and that regime is exactly the one that pushes hyperscalers toward building cheaper chips in-house. Amazon's triple-digit chip growth is the tell. The constructive lean stays on the in-house-silicon franchises (Google's TPU, Amazon's Trainium) and the picks-and-shovels around them (networking, connectivity), with caution on the multiple of everything downstream of a buildout that now has to pay for itself at 5%-plus long-term rates.

Stocks in play

  • AMZN (Amazon), the week's main custom-silicon event. Bull: AWS +37% (fastest in 18 quarters), AWS AI business above $25B run-rate growing triple-digits, and, new and specific, the homegrown chip business (Trainium, Graviton) "growing triple digit percentages," with Jassy "enthusiastic about the ROIC equation." Bear: capex hiked to $220B, free cash flow -$7.6B over the trailing year, long-term debt doubled to $129B, and its two biggest AWS customers are Anthropic and OpenAI. Watch: whether Amazon ever breaks out Trainium revenue, and a 2027 capex number (not yet given).
  • MSFT (Microsoft). Bull: Azure +43% (fastest since 2022), above $100B annual run-rate, ~45% guided next quarter; Copilot 30M paid seats (+50% in three months); capex held flat at ~$175B, the market's model for "disciplined spending that works." Bear: the custom-silicon story at Microsoft remains a black box, and growth leans heavily on OpenAI. Watch: any Maia 2 mention and whether free cash flow turns clearly positive.
  • META. Bull: core ad business still humming (+28% revenue); Zuckerberg is at least talking about selling compute and "selling intelligence" at higher margins. Bear: free cash flow collapsed to $784M (from $8.6B), capex raised with no monetization story, Reality Labs still bleeding ($4.62B quarterly loss), stock -10%. Watch: any MTIA (Meta's in-house inference chip) deployment detail, and whether the "sell compute" idea gets a customer or a timeline.
  • NVDA (Nvidia). Bull: crossed $5 trillion in market cap; NVLink still owns ~95% of scale-up networking; keeps buying into the inference threat (the Groq licensing deal); the cloud acceleration everywhere is raw GPU demand. Bear: the CUDA moat is showing hairline cracks as open-weight models ("open doesn't need CUDA") take routing share, and a regime that punishes capex hits the highest-margin link, merchant GPUs, first. Watch: whether open-weight share keeps climbing, and any official read on Rubin timing.
  • GOOGL (Alphabet), recap only. Bull: last week's blowout still frames the sector, Google Cloud +82% was the number every other hyperscaler was measured against this week, and the TPU is the ASIC that matters. Bear: first-ever negative free cash flow and ~$200B capex still hang over it. Watch: any external-TPU disclosure and next quarter's cash flow.
  • AVGO (Broadcom) / MRVL (Marvell) / ALAB (Astera Labs), read-through. Bull: the networking-market framing (ESUN, above $100B by 2030, Broadcom among the 12 founding ESUN members) is a direct tailwind to the merchant connectivity and custom-XPU arms dealers as hyperscaler chips scale. Watch: a design-win disclosure tied to Trainium 3 / TPU v7 / Maia 2, and any read on switch-chip competition from new entrants like Upscale.
  • ARM. Arm CEO Rene Haas was interviewed on Squawk on the Street on 7/30, with results "driven higher" by "a surge in AI demand." Watch: the full Haas interview for royalty-mix color on Neoverse in hyperscaler ASICs.
  • Private and adjacent: Etched, Groq, Cerebras, Upscale. The inference-ASIC and AI-networking startup wave is where this week's genuinely new custom-silicon energy was: Etched ($300M, ~$10B), a 10-plus-company field, and Upscale's "Skyhammer" scale-up switch chip working with "a few hyperscalers." Watch: tape-out timing against the capex cycle, which the bulls themselves called the key risk.

Read-throughs

  • Networking is the read-through nobody's pricing. Every custom accelerator needs a fabric to talk to its neighbors, and that layer is becoming a market of its own: more than $100B cumulative by 2030 per Dell'Oro, with an open ESUN standard forming to challenge Nvidia's NVLink (~95% share today) and 175 companies already signed up. Names: Broadcom (Tomahawk switches, ESUN founding member), Marvell (Teralynx), Astera Labs (connectivity attach), plus insurgents like Upscale (Skyhammer). If custom chips scale, the connectivity attach scales with them regardless of who wins the logic war.
  • Memory is the toll booth, and it's getting more expensive. Multiple hyperscalers blamed higher memory costs for lifting capex this week (Amazon explicitly tied its $220B number to memory), and Micron crossed a $1 trillion market cap. HBM (Hynix, Micron, Samsung) gets paid on every accelerator, custom or merchant, and rising memory prices raise the cost of everyone's chip, which paradoxically strengthens the case for squeezing cost elsewhere via custom logic.
  • TSMC and advanced packaging. Advanced packaging remains the shared choke point behind both GPUs and ASICs. Watch for it to re-emerge as the 2027-28 designs (TPU v7, Trainium 3, Maia 2, the "Frozen" branch) actually ramp.
  • The financing plumbing, the fault line under everything. Yields at 2007 highs, a barely-oversubscribed hyperscaler bond, Meta's free cash flow near zero, Amazon negative $7.6B. This is the shared risk to every name in the newsletter. Dearer money favors cheaper (custom) silicon at the margin, but if it chokes off the buyers entirely, the whole chip debate goes on hold.

What changed vs last week

Last week (issue dated 2026-07-27) was Alphabet's week: Google Cloud +82%, the TPU turning into an external business, the "Frozen" next-gen chip leaking, Nvidia's Vera Rubin specs landing, and the first airing of the AI-financing bear case. This week three things concretely changed:

  1. The capex reckoning went from one company to the whole group, and split it. Last week Alphabet alone got punished for spending. This week the market applied the new rule to everyone at once and graded on a curve: Microsoft rewarded for holding capex flat and proving monetization (Azure +43%, Copilot 30M), Meta hammered for raising capex with no payoff story (FCF $784M, -10%), Amazon given a pass for hiking to $220B because AWS and its own chips are visibly working. The narrative moved from "is AI spending good?" to "whose AI spending is good?"
  2. Amazon's in-house silicon got its first real quantified callout. Last week the custom-silicon proof point was Google's TPU. This week Amazon supplied the sequel: Trainium and Graviton "growing triple digit percentages," alongside an AWS AI run-rate above $25B. Two of the four big in-house programs (TPU, Trainium) now have real momentum behind them in as many weeks.
  3. The startup and networking layers came into view. Last week was all hyperscaler prints. This week the ecosystem showed up: an inference-chip startup (Etched, $300M) into a 10-plus-company field, and a serious case that AI networking (ESUN, Upscale's Skyhammer) is the next hundred-billion-dollar market.