Newsletter · · Ashutosh Agarwal

Most Custom AI Chips Will Fail, Top Investors Warn - Custom Silicon vs Nvidia - Week of July 20, 2026

Custom-silicon podcast intelligence for the week of July 13 to 20, 2026. Top investors including Gavin Baker argued most ASIC programs will fail and only Google's TPU truly threatens Nvidia, while the memory trade reversed and Jim Chanos put hard numbers on the bear case.

Custom Silicon vs Nvidia

Week of July 20, 2026: Most Custom AI Chips Will Fail, Top Investors Warn

Custom Silicon vs Nvidia, weekly, issue dated 2026-07-20.

A note before we start: Last week the ASIC bulls got their trophy, Apple signed a real, dated Broadcom design win. This week the pendulum swung the other way, and it swung on the strength of who was talking. No new design win landed. Instead, two of the most-listened-to investors in the game sat down and said, in effect, slow down, most of these custom-chip programs are going to fail, and the one that actually threatens Nvidia is Google's, not the long list of hopefuls. At the same time, the memory trade that detonated into last week's headline went into reverse: SK Hynix, the stock everyone piled into after its record US listing, posted its biggest one-day drop ever. So this was a debate-and-reality-check week, light on scoreboard, heavy on the smartest people in the room telling you which parts of the story are real. Worth reading closely, because sentiment is what's moving these stocks right now, and sentiment just turned.

One housekeeping note on sourcing: almost everything actionable this week came from pundits and investors, not from operators or insiders. There was no hyperscaler chip boss on camera.

TL;DR

  • "Most of those ASICs are gonna fail." That's Gavin Baker, CIO of Atreides Management, one of the most respected tech investors alive, on The a16z Show. His forecast: "In the next 3 years, I think you'll see a bunch of high-profile ASIC programs canceled," especially if Google starts selling its TPU chips on the open market. In his telling, this isn't ten companies versus Nvidia. It's really just Google's TPU (built with Broadcom, for now) versus Nvidia, with Amazon's Trainium a distant, improving third and AMD as the perennial "second source." [PUNDIT]

  • The other big a16z conversation reached the same place from a different door. An infrastructure expert on The a16z Show called custom silicon "the biggest thing" threatening Nvidia, Amazon making "millions of Trainium," Google making "millions of TPUs" running at "100% utilization", but then laid out the brutal math for why almost nobody can win: to beat Nvidia you need a 5x hardware edge, and Nvidia's supply-chain scale grinds that 5x down to "50% better," at which point it's not worth it. Microsoft's custom silicon, he said flatly, "kind of sucks." [PUNDIT/EXPERT]

  • The memory mania reversed. SK Hynix, last week's darling after the largest foreign IPO in US history, "suffers its biggest one-day drop on record," per Bloomberg Tech, a ~15% crater in Korea that dragged Samsung, the Kospi, and US-listed AI chip names down with it. The verdict from the desk: "buy the rumor, sell the fact." The ~$4.4 trillion trio of TSMC, Samsung, and SK Hynix is suddenly seeing "the shine wear off." [NEWS]

  • Nvidia's roadmap got the rumor treatment, and the semis pros shrugged. The Circuit walked through a pile of leaks: Rubin Ultra "scaled down from four die to two," the Kyber rack delayed, co-packaged optics slipping (Nvidia going "NPO instead of CPO"). Their read: none of it dents Nvidia's revenue given demand, but it matters a lot to the small optical and PCB names one layer down. [PUNDIT/EXPERT]

  • The bear got specific. Jim Chanos, on RiskReversal, put numbers on his skepticism: hyperscaler return on incremental invested capital has fallen "from... 40% a year and a half ago to about 20% today," heading "toward 10%." His rule of thumb: "No one that is dependent upon NVIDIA to exist should trade at a higher valuation than NVIDIA itself. And yet many, many companies do." [PUNDIT]

What's new

Ranked by what actually moves a book. Two caveats up front: (1) there was no fresh design-win leak this week, the top of the list is high-quality investor opinion, not a signed contract; (2) I've separated the genuinely new items from the recaps.

1. Gavin Baker's ASIC cull: "most of those ASICs are gonna fail." [PUNDIT, named buy-side] The single most thesis-relevant thing said all week. On The a16z Show (2026-07-14), Gavin Baker, Managing Partner and CIO of Atreides Management, and the guy half of tech Twitter reads to figure out "what the F is really going on", was blunt: "I personally believe most of those ASICs are gonna fail... In the next 3 years, I think you'll see a bunch of high-profile ASIC programs canceled, especially if Google starts selling TPUs externally, which has been all over X." His framing collapses the crowded ASIC field into a two-horse race: "it's really a battle between Google and its TPU enabled by Broadcom for now. And Google can take the TPU away from Broadcom whenever they want. Now they can't do the Ethernet networking that Broadcom is doing." On the rest of the pack he was more nuanced than dismissive, Amazon's Annapurna group is "arguably the most talented silicon team at any hyperscaler," and "the Trainium 3 will probably be a much better chip than the Trainium 2" (with the sober reminder that "it took Google 3 generations to get the TPU right"). And he thinks AMD survives as the necessary runner-up: "AMD will always be kind of the second source and you need a second source." Why it moves a book: this is a direct rebuttal to last week's ASIC-bull momentum from the most credible possible source. If Baker is right, the trade isn't "own the basket of custom-silicon winners", it's "own Google's TPU exposure and Broadcom's cut of it, and be very careful about everyone else."

He also handed the bulls their macro anchor. Asked if we're in an AI bubble, Baker said no, and explained why with a comparison worth writing down: at the 2000 peak, "97% of the fiber that had been laid in America was dark. Contrast that with today. There are no dark GPUs." Cisco peaked at "150 or 180 times trailing earnings. Nvidia's at more like 40 times." And the hyperscalers spending the money have seen "call it a 10-point increase in their ROICs" since ramping capex. His stage-setting stats: ~"$1 trillion of data centers in the US," a plan to add "$3 to $4 trillion in the next 5 years," Google processing "150x" more tokens than 17 months ago, and OpenAI sitting on "$1 trillion of deals."

2. The infrastructure expert's "5x rule": why beating Nvidia is nearly impossible. [PUNDIT/EXPERT] The companion a16z episode, "Can Anyone Catch NVIDIA?" (2026-07-15), is the best plain-English explanation of the whole debate I've heard this year. Asked how threatened Nvidia is by custom silicon, the lead speaker (a deep supply-chain infrastructure expert) said: "I think that's the biggest thing... Amazon is making millions of Trainium, Google's making millions of TPUs. TPUs clearly are like 100% utilized... Trainium's not there, but I think Amazon will figure out how to do that." The hyperscalers, he said, are "kind of lucky", they can copy Nvidia's approach and win purely on supply chain, because they have a captive customer (themselves). "It's a margin compression exercise, essentially." Meta might specialize for recommendation systems; Microsoft's custom silicon, in his words, "kind of sucks."

But for anyone without a captive customer, he laid out why the wall is so high. "To beat NVIDIA... you need to do something that will give you 5x advantage in hardware efficiency for a certain type of workload. And then pray the workload doesn't shift." And even a 5x design edge gets ground down: "the supply chain stuff means that 5x actually turns into a 2.5x. And then NVIDIA can compress their margin a little bit... And then that 2.5x becomes like a 50% better." His exhibit A is AMD: "They got to 2 nanometer before NVIDIA. They had higher density HBM. They use 3D stacking... and yet they still lose." The economics behind it: Nvidia runs a "75% gross margin," AMD "50%." He also flagged the froth in merchant challengers, Etched and Rivos have "raised billions" without ever launching a chip publicly. On Google, though, he agreed with Baker's provocation: "I totally think Google should sell TPUs externally, not just renting, but like physically", noting Google is "even discussing it internally," but it would require "a big reorg of culture." Actionable read: this is the analytical case for staying long Nvidia and long the arms-dealer (Broadcom) rather than betting on the long tail of accelerator startups.

3. Nvidia's roadmap wobble, and why the pros say it doesn't matter (for Nvidia). [PUNDIT/EXPERT] The most concrete new datapoint on the merchant side. On The Circuit (2026-07-14), the semis-focused hosts catalogued a week of Nvidia supply-chain rumors: "Ruben ultra was getting scaled down from four die to two. It was Kyber rack was delayed. It was the copac optics is getting delayed. They're going to NPO instead of CPO" (near-package optics instead of the more ambitious co-packaged optics). Nvidia's official posture, they noted, is "no delays. Our timeline remains intact." Their own take is the useful part: even a "six months at a worst case scenario" delay "does not change the revenue structure... there's so much demand for compute." The real damage, they argued, lands on the layers below Nvidia: "for that part of the supply chain, missing a quarter or not being recognized revenue in that chip corner, matters more than it does for NVIDIA." In other words, treat Rubin-timing headlines as a supplier problem (optical, PCB, packaging names), not an Nvidia problem, and expect "the street... arguing about this stuff for the next 18 months."

4. The memory trade goes into reverse. [NEWS] Last week's headline star became this week's cautionary tale. On Bloomberg Tech (2026-07-14), the open said it plainly: "an AI-fueled stock rout in South Korea spills over into the US market... SK Hynix suffers its biggest one-day drop on record", a roughly 15% fall in Korea that hit Samsung and the broader AI chip complex, in just the second US trading session after Hynix's record-breaking debut. The ADRs, priced at "$149," were trading "just under $158" (down about 6% that session, after opening near $170 the week before). Swissquote's analyst chalked it up to "the very usual investor behavior of buy the rumor, sell the fact," warning that "downside risks do prevail as long as the volatility remains this high", with leveraged ETFs driving "double-digit percentage up and downs... every single day almost." Bloomberg's Dina Bass kept the structural bull case honest: the memory makers are genuinely "trying to leverage the interest in AI to make their market less cyclical" via multi-year customer deals, and don't "expect supply to catch up with demand any time soon." But she also named the long-term risk nobody wanted to discuss last week: if HBM is such a chokepoint, there's "a strong incentive to figure out a way to do it differently or to use less of it." The macro tell: the ~"$4.4 trillion" combined value of TSMC, Samsung, and SK Hynix is now something big EM funds are "looking to rotate out of."

5. Chanos puts numbers on the bear case. [PUNDIT] On RiskReversal (2026-07-17), short-seller Jim Chanos delivered the most quantified skepticism of the week. His core objection is a duration mismatch, "people are committing long-term capital projects based on near-term spot pricing", a 20-year data-center asset underwritten on one-to-two-year rental commitments, which he compared to shale, 19th-century railroads, and the GFC repo market. The metric he says he's watching for clients: hyperscaler return on incremental invested capital has fallen "from... 40% a year and a half ago to about 20% today," and "if the spend keeps up at this kind of rate, it's going to be moving toward 10%." At that point, he argues, "you're going to have real issues at the hyperscaler C-suite." On Micron specifically, directly relevant to the HBM trade, he was scathing: "a company that had negative gross margins three years ago," whose stock went from "$100... 18 months ago" to a top of "$1,250 just recently," now "$950." His valuation rule for the whole ecosystem: "No one that is dependent upon NVIDIA to exist should trade at a higher valuation than NVIDIA itself. And yet many, many companies do." He read Nebius's new "asset-light" pivot, from the most capital-heavy neocloud, which spent "$4.40 in capital to generate $1 in revenue", as "a tremendous admission." Why it matters for our beat: Chanos is not shorting the chips, he's shorting the financing structure underneath the buildout, and that's the risk that would hit ASIC and GPU demand alike.

6. The Nvidia bull's rebuttal: CUDA, $110B of hoarded supply, and a "mean reversion" call. [PUNDIT] For balance, on Monetary Matters (2026-07-14), investor Ben Pouladian made the case that "Nvidia is mispriced" to the upside. Two facts worth stealing: Nvidia "spent $110 billion on their supply chain, basically buying everything up" to make sure a single missing "screw or one cable wire" can't delay a data center, and as a result "there's not an actual GPU shortage anymore," yet pricing power holds because of hardware-software co-design. His most tradeable observation is a relative-value one: "I don't think Nvidia can be at like such a low multiple and then semi-cap [ASML, Lam, Applied Materials, KLA], which has always been the low multiple guy, be at such a high multiple. So there must be some sort of mean reversion at some point." On the CUDA moat, he made the underappreciated point that Nvidia's open-source software (like Nemotron) is "optimized on CUDA," so developers worldwide improve the ecosystem "for free", a flywheel a fresh ASIC doesn't have.

7. Meta's compute-rental economics get real, and Apple goes to court. [NEWS, recaps and adjacent] Two threads that touch our beat without being about chips directly. First, per the Elon Musk Podcast (2026-07-19, a news-recap show, treat as reported news, not primary), Meta is "in early talks to lease some of its data centers to Anthropic in a deal worth up to $10 billion over two years," while Anthropic separately signed a deal "reportedly worth $35 billion over a three-year period" with Elon Musk's xAI, "$15 billion a year purely to rent computing hardware." Meta expects to spend "$125 billion and $145 billion a single year" on data centers, and Zuckerberg "fields requests to buy data center space almost every week." This is the demand-side backdrop that makes Meta's own Iris chip (production this September, a recap from last week, re-noted on Everyday AI [2026-07-13], where it was framed as supplementing "rather than replacing NVIDIA's and AMD's GPUs") economically important: a cheaper in-house chip fattens the margin on every rented rack. Second, Everyday AI and 20VC (2026-07-16) both covered Apple suing OpenAI on Friday, alleging its consumer-hardware team stole trade secrets via two ex-Apple employees. Not a silicon story, but it deepens the Apple-vs-OpenAI rift right as both race for their own chips and TSMC allocation. Also from Everyday AI: Microsoft has quietly started routing "some Excel and Outlook features" to its own internal MAI models, with the honest caveat that "Microsoft's own models aren't very good yet," which rhymes with the a16z expert's "their custom silicon kind of sucks."

The debate

For ASICs (they take share and cap Nvidia's margins). The strongest version of this case is no longer "everyone builds a chip." It's narrower and, honestly, more convincing for it: Google's TPU is a genuine, at-scale, ~100%-utilized franchise that took three generations to get right and is now good enough that Baker thinks Google could sell it "physically" on the open market, a move "all over X" and reportedly what Anthropic wants ("tens of billions of TPUs"). Amazon's Annapurna team is the real deal, and Trainium 3 should be a step-change. The hyperscalers don't need to beat Nvidia on engineering; they win by margin compression against a captive internal customer. And the demand is so vast ($3–4 trillion of data centers coming) that even a modest share shift is enormous dollars.

For the merchant (Nvidia keeps the pie). This side had the louder week. Two independent, credible voices converged on the same conclusion: the long tail of ASICs mostly fails. The 5x rule is why, a design edge gets ground to "50% better" once Nvidia's supply-chain scale and 75% gross margin come into play, and by then it isn't worth the multi-year effort while the workload shifts under you. AMD is the proof: better process node, better HBM density, 3D stacking, and it still loses. Nvidia has no GPU shortage (it bought $110B of supply insurance), still has CUDA's free global developer flywheel, and even its "roadmap slip" is dismissed by the semis pros as a supplier problem, not a demand problem. And the memory reversal is a useful reminder: the AI-infrastructure trade can reprice violently on sentiment without any change in the underlying franchise.

The wildcard both sides now agree on: Google. Notice that the bull and bear framings this week both funnel to the same name. The threat to Nvidia isn't the basket, it's Google specifically, and specifically if Google decides to sell TPUs as merchant silicon. That is the binary to watch.

Where I come out this week: the ASIC story didn't advance, it got pruned. The smartest voices took a crowded, hype-y field and cut it down to "Google's TPU, Amazon's Trainium as a maybe, and a graveyard of the rest." That's actually bullish for the two names that get paid regardless, Nvidia (whose moat just got re-argued by people with no reason to shill it) and Broadcom (Google's current TPU partner and the one arms dealer sitting on the right side of the one franchise that matters). The genuine new risk isn't a competing chip; it's Chanos's financing math. Watch ROIIC, not benchmarks.

Stocks in play

  • NVDA (Nvidia). [PUNDIT week] Bull: the moat got its best independent defense of the year, the 5x rule, 75% gross margins, $110B of pre-bought supply, no GPU shortage, and CUDA's free developer flywheel (The a16z Show, Monetary Matters). Roadmap rumors are a supplier problem, not a revenue problem (The Circuit). Bear: Rubin Ultra reportedly cut from four die to two and optics slipping to NPO, if demand ever softens, those slips start to bite; and Chanos's ROIIC math is an existential-if-slow risk to the buyers (RiskReversal). Watch: whether the four-die-to-two Rubin Ultra rumor gets confirmed or denied, and the next hyperscaler capex-vs-operating-income prints.

  • GOOGL (Google/Alphabet). [PUNDIT] Bull: this week crowned the TPU as the only custom-silicon program the smart money genuinely fears, ~100% utilized, three generations mature, and a potential external-sales option that Baker thinks could be worth more than people realize (The a16z Show). Bear: selling TPUs externally would require "a big reorg of culture," and Google itself still values Gemini far above the chip business; nothing is committed. Watch: any concrete move toward selling TPUs as merchant silicon, and the rumored Anthropic TPU purchase.

  • AVGO (Broadcom). [PUNDIT, read-through] Bull: if the field narrows to "Google's TPU vs Nvidia," Broadcom is the arms dealer on the winning custom side, it builds the TPU "for now" and owns the Ethernet networking Google can't easily replace (The a16z Show). Bear: Baker's own caveat, "Google can take the TPU away from Broadcom whenever they want", plus Chanos's rule that names dependent on the buildout shouldn't out-trade Nvidia. Watch: whether Google keeps Broadcom on future TPU generations, and OpenAI/other XPU ramps.

  • AMZN (Amazon). [PUNDIT] Bull: the Annapurna team is "arguably the most talented silicon team at any hyperscaler," and Trainium 3 "will probably be a much better chip than the Trainium 2" (The a16z Show). Bear: Trainium is "not there" on utilization yet, and getting a chip right historically takes three generations. Watch: Trainium 3 performance disclosures and any external-sales traction.

  • META. [NEWS] Bull: the compute-rental business is getting real, up to $10B from Anthropic, on top of $125–145B of annual data-center spend, and the Iris chip (production in September) fattens the rental margin (Elon Musk Podcast, Everyday AI). Bear: Iris is a first production run after a long in-house struggle, and the a16z expert singled Meta's custom silicon out only as a possible winner for recommendation workloads, not general AI. Watch: the next capex number and any Iris deployment scale.

  • MSFT (Microsoft). [PUNDIT] Bull: still in the game on both models (now routing some Excel/Outlook features to in-house MAI models) and Maia silicon. Bear: two independent unflattering data points this week, "their custom silicon kind of sucks" and "Microsoft's own models aren't very good yet" (The a16z Show, Everyday AI). Watch: any Maia 2 timing or benchmark that contradicts the skeptics.

  • AMD. [PUNDIT] Bull: the necessary "second source", "you need a second source," and buyers will always route some volume to a viable #2 (The a16z Show). Bear: the cleanest cautionary tale in the whole debate, better node, better HBM, 3D stacking, and it "still loses" to Nvidia, with margins half Nvidia's (The a16z Show). Watch: whether Microsoft/Meta re-accelerate AMD purchases.

  • Memory, SK Hynix / Samsung / MU (Micron). [NEWS + PUNDIT] Bull: still the scarcity story; makers are locking multi-year deals to break the cycle and don't see supply catching demand (Bloomberg Tech). Bear: the reversal is the story, SK Hynix's biggest one-day drop on record, "buy the rumor sell the fact," leveraged-ETF whipsaw, and Chanos's Micron takedown (negative gross margins three years ago, stock $100 → $1,250 → $950) (Bloomberg Tech, RiskReversal). Plus the new long-term worry: a strong incentive across the industry to "use less" HBM. Watch: whether the multi-year contracts actually hold through a sentiment drawdown.

  • MRVL / ALAB / ARM / Alchip. [COVERAGE GAP] No fresh Marvell ASIC-roadmap, Arm/Neoverse royalty, or Alchip color this week. Astera Labs was flagged "in focus" on Stock Market Today With IBD (2026-07-15) as a connectivity name competing in AI data centers, but with no new attach-rate detail. Second quiet week in a row for the connectivity/IP layer, expect it to re-surface as the '27 designs (TPU v7, Trainium 3, Iris, OpenAI/Apple silicon) actually ramp.

Read-throughs

  • TSMC / advanced packaging, still the choke point under everything. Bloomberg Tech pegged TSMC June sales up "68% year on year," with a Bloomberg calc that June-quarter sales "jumped 36%," and earnings due that Thursday (2026-07-14); The Financial Exchange (2026-07-16) built its whole "AI chip volatility" segment around the TSMC print and US fab investment. The primer point that matters for our beat: TSMC does "over 90% of sub-7nm chips" and leads advanced packaging (Chit Chat Stocks), so every GPU and every ASIC funnels through the same door. Whoever wins the logic war, TSMC gets paid.

  • HBM, the bottleneck that could become a liability. Last week HBM was the toll booth; this week Bloomberg Tech surfaced the flip side, if HBM stays this scarce, there's "a strong incentive to figure out a way to do it differently or to use less of it," which over time is a risk to the memory-maker dominance, not just a tailwind (2026-07-14). And Chanos's Micron history lesson is a reminder that memory has always been cyclical until proven otherwise (RiskReversal).

  • Networking, "your GPUs are waiting for something." AI Proving Ground (2026-07-15) made the case that the bottleneck inside a cluster is increasingly the network, walking through Ethernet vs InfiniBand and the switching-ASIC layer (Spectrum, Cisco Silicon One). It stopped short of NVLink/UALink/Tomahawk/Teralynx specifics, but the direction of travel is the read-through: connectivity content per accelerator keeps rising, which is the structural bull case for Broadcom's and Marvell's switching franchises and for Astera Labs' attach, even in a week when none of them made fresh news.

  • The financing plumbing, the real fault line. Chanos's ROIIC-decline chart (hyperscaler incremental returns 40% → 20% → toward 10%), the Nebius asset-light pivot, and Meta's rush to rent out spare compute all point at the same thing: the money underwriting this buildout is getting thinner per dollar spent (RiskReversal, Elon Musk Podcast). That's the shared risk to all accelerators, merchant and custom alike, if the buyers blink, the chip debate becomes academic.

What changed vs last week

Last week (dated July 13) the story hardened into facts and rotated into memory: Apple became a signed Broadcom design win, Meta stamped a September production month on Iris, Nvidia got repriced to ~18x, and the memory complex detonated into the headline on SK Hynix's record IPO. This week the mood flipped in three specific ways:

  • The ASIC narrative reversed from "the cast list keeps growing" to "most of them will fail." Last week's momentum was additive, another rumored buyer, another program. This week the two most credible voices on the topic (Gavin Baker; the a16z infra expert) argued the opposite: the field is too crowded, the 5x wall too high, and the real contest narrows to Google's TPU vs Nvidia. That's a genuine sentiment turn on the core thesis.

  • Memory went from mania to reversal. Last week: largest foreign IPO in US history, Samsung out-earning Nvidia, a chairman saying doubling capacity isn't enough. This week: SK Hynix's biggest one-day drop on record, "buy the rumor sell the fact," leveraged-ETF whipsaw, and big funds rotating out of the ~$4.4T TSMC/Samsung/Hynix trio. Same names, opposite tape.

  • The bear case got specific and numeric. Last week the risk was framed as "existential but 2–3 years out." This week Chanos gave it a metric to watch, hyperscaler ROIIC falling toward 10%, and a valuation rule ("nothing dependent on Nvidia should trade above Nvidia") that indicts large parts of the AI-infrastructure complex today, not in 2028.

What did not change: still no published ASIC benchmark beating Blackwell, still no hard shipment/volume numbers for Maia, MTIA/Iris, or Trainium, and still a quiet week for the connectivity/IP layer (Marvell, Astera Labs, Arm, Alchip). The debate got sharper; the scoreboard still hasn't moved.