Newsletter · · Ashutosh Agarwal

Google Starts Selling TPUs Externally as the AI Capex Reckoning Arrives - Custom Silicon vs Nvidia - Week of July 27, 2026

Google's Q2 print turned the TPU into an external, paying business and pushed Alphabet to its first-ever negative free cash flow, alongside the leaked Frozen chip, Nvidia's Vera Rubin specs, and a widening capex reckoning, for the week of July 27, 2026.

Custom Silicon vs Nvidia

Week of July 27, 2026: Google Starts Selling TPUs Externally as the AI Capex Reckoning Arrives

A note before we start: Last week the smartest voices in the room argued about custom silicon, Gavin Baker said most ASIC programs would fail, the a16z crowd explained why beating Nvidia is nearly impossible, and Jim Chanos put numbers on the bear case. This week the arguing stopped and the facts arrived. Alphabet reported after the bell on Wednesday, and buried in a "monster quarter" were the two things this newsletter exists to track: Google's homegrown TPU chip quietly turned into a real, external, paying business for the first time, and a report leaked that Google's next chip, codenamed "Frozen", is already in the works. Then the market did something it hasn't done in two years: it punished Alphabet for spending on AI, sending its stock down ~7% and its free cash flow negative for the first time in the company's history. So the two threads finally braided together this week, the custom-silicon story got real, and the financing-risk story that Chanos flagged last week went from thesis to headline.

One sourcing note. This was a heavy earnings-and-macro week, so a lot of the color comes from broadcast shows reacting to prints (CNBC's Squawk on the Street, The Exchange) and from investor podcasts, plus one genuinely useful semis deep-dive. There was no hyperscaler chip boss on a microphone. But we did get something better than opinion this week: actual numbers. I've tagged every item [OPERATOR], [NEWS/PRINT], or [PUNDIT] so you know exactly what you're getting.

TL;DR

  • The TPU became a revenue line. On Alphabet's Q2 call, Google Cloud grew 82% to $24.8 billion (Wall Street wanted 63%), and the reason the Street kept naming was custom silicon plus Anthropic. As one host put it on The Rundown, the backlog is now $514 billion. And per reporting relayed on Tech Brew Ride Home and Limitless, Google in Q2 shifted from using TPUs mostly internally to selling them to outside AI labs, exactly the "sell TPUs externally" move Baker flagged last week as the real threat to Nvidia. [NEWS/PRINT]

  • Meet "Frozen." The Information's reporting (relayed on Tech Brew Ride Home, 2026-07-20) says Google engineers project a new chip, codenamed Frozen v2, could be 6–10x more efficient than today's TPU on tokens-per-watt, shipping as soon as 2028. The twist: it bakes Gemini's model decisions into the silicon ("freezing" them), a bet that Google keeps its current model architecture. It's a new branch, not a TPU replacement. [NEWS/PRINT]

  • Nvidia's answer got loud. On Limitless (2026-07-24), the hosts claimed Nvidia's Vera Rubin platform is now standing up: 10x more tokens per watt than its predecessor, 336 billion transistors (+62% vs the Blackwell GB300), on TSMC 3nm. Treat the "it's shipping today" framing as excited pundit talk, but the spec gap is the point: even a 6–10x Google chip in 2028 has to beat whatever Nvidia ships by 2028. [PUNDIT]

  • The capex mood snapped. Alphabet raised full-year 2026 capex to $195–205 billion (from $180–190B), spent ~$45B in the quarter, and printed negative $5.9 billion free cash flow, its first-ever negative FCF quarter. Steve Eisman's read on The Real Eisman Playbook: "the terms of debate on AI have truly shifted." The stock fell ~7%; the Nasdaq dropped >2% Thursday. [NEWS/PRINT]

  • The financing plumbing is now the story. On Big Technology, the hosts walked through Ed Zitron's math: Bloomberg estimates >$500 billion of AI data-center debt (≥$200B in private credit, ~8% of all private-credit loans), and roughly 70% of hyperscaler capacity is committed to OpenAI and Anthropic, "dual points of failure." This is the risk that hits every accelerator, custom and merchant alike. [PUNDIT]

What's new

Ranked by what actually moves a book: real prints and reported design/monetization news first, pundit hot takes last.

1. Google's TPU crossed from cost center to profit center, and the Street noticed. [NEWS/PRINT] This is the single most thesis-relevant development of the week, and it came straight from Alphabet's Q2 earnings (reported after the bell Wednesday 7/22). The topline was a blowout: revenue of nearly $120 billion, up 24%, with Google Cloud up 82% to $24.8 billion, the second straight quarter of acceleration (prior quarter was +63%), against a Street expectation of just 63% (The Rundown, 2026-07-23; The Real Eisman Playbook, 2026-07-24). The remaining performance obligation, the contracted-but-not-yet-collected "backlog", is now $514 billion, and Google says more than half converts to revenue within two years (The Rundown).

Why it belongs at the top of a custom-silicon newsletter: the growth is being driven by AI infrastructure that runs substantially on Google's own TPUs, and, new this quarter, Google started selling that silicon to outsiders. As the Limitless hosts summarized it: "they made quite a decent chunk of money selling their TPUs to external customers... a big jump from the previous quarter, because Google primarily built and used TPUs to train their own Gemini models. What they've done, the change between Q1 and Q2, is they've started selling it to other vendors, such as Anthropic and other AI labs" (Limitless, 2026-07-24). On CNBC, Jim Cramer put it plainly: "TPU, their own semis were really good. They have a lot more business than they can handle" (Squawk on the Street, 9am, 2026-07-23). Remember what Baker said last week, the real danger to Nvidia isn't the crowded ASIC field, it's Google specifically deciding to sell TPUs on the open market. This quarter, that started happening.

The nuance worth holding onto: a large chunk of this is one customer. An analyst on Squawk estimated that roughly two-thirds of Alphabet's backlog growth over the past year is Anthropic (Squawk on the Street, 10AM, 2026-07-22). So "Google's TPU is now an external business" is really, for now, "Anthropic is renting a lot of TPUs." Concentration cuts both ways.

2. "Frozen v2": Google's next chip leaks, and it's a bet on baking the model into the silicon. [NEWS/PRINT] The concrete design news of the week. Per The Information's reporting (read out in detail on Tech Brew Ride Home, 2026-07-20), Google is building a new chip codenamed "Frozen," with engineers projecting it "could be 10 to 6 times more efficient than the newest version of Google's existing line of homegrown AI chips when it launches, based on the number of tokens... it can serve per unit of power" (Tech Brew Ride Home, 2026-07-20). The name is the whole idea: "The 'Frozen' name comes from the idea of permanently etching parts of the model into the silicon."

Here's the part that matters for the debate. Today's TPUs, like Nvidia's GPUs, are general, "made to work with many different AI models... those chips must make a lot of time-consuming decisions as they interact with whichever model they're running." Frozen v2 would instead have "some of the decisions for Google's Gemini models built in, reducing the number of steps the chip must make... and the amount of data it has to move around" (Tech Brew Ride Home). In plain English: it trades flexibility for speed and efficiency by assuming Google will keep building Gemini the same way. That's a real bet, if the model architecture changes, the chip's edge erodes. It's also explicitly a new branch of silicon, not a TPU replacement, and Google plans to deploy it "as soon as 2028."

Google isn't alone in chasing inference-specialized silicon: the same report names SambaNova, d-Matrix, OpenAI, and Microsoft building similar inference chips, all aimed at "handling inference more efficiently than Nvidia's GPUs to alleviate shortages of computing power that have driven up costs." And a striking tell about how seriously Nvidia takes the inference threat: "Nvidia itself struck a $20 billion deal in December to license technology from Groq, a startup that designs inference chips" (Tech Brew Ride Home). The Limitless hosts framed Frozen more skeptically, 6–8x efficiency but "isn't coming online until 2028," by which point Nvidia will have moved on (Limitless, 2026-07-24). Both things can be true: a real, funded next-gen program, and a long road to relevance.

3. Nvidia's Vera Rubin spec sheet lands, and it resets the bar. [PUNDIT] For the merchant side, the Limitless hosts walked through Nvidia's Vera Rubin platform, which they described as newly operational (I'd treat the "shipping today" enthusiasm as pundit excitement, not a confirmed operator disclosure, but the specs are the useful part). Their numbers: 10x the tokens per watt of the prior generation; 336 billion transistors, "a 62% increase over the Blackwell GB300"; built on TSMC's 3nm; 22 terabytes per second of memory; two maxed-out dies "glued" into one; roughly 4x the baseline performance of the GB300; and a rack architecture of 72 GPUs serviced by 32 CPUs (Limitless, 2026-07-24). They also floated an eye-watering, back-of-envelope revenue extrapolation, "1,000 racks per day... $630 billion per quarter... assuming they can get this up at scale", which is pure pundit math and should be treated as such, not a forecast.

Why it matters for our beat: the 10x-per-watt claim is aimed squarely at the exact metric Google's Frozen chip is chasing (tokens per unit of power). The Limitless takeaway, "even if they achieve six to eight times more efficiency in two and a half years' time, who the hell cares? Because Nvidia will have a better chip architecture by then", is the merchant bull case in one sentence: it's not enough for a custom chip to beat today's Nvidia; it has to beat 2028's Nvidia. That's the treadmill.

4. The Ben Thompson framing: Nvidia's margins are a "price umbrella," not a moat. [PUNDIT] The most useful analytical piece of the week came via Tech Brew Ride Home reading a Ben Thompson (Stratechery) essay. His argument cuts against the "tokens are a commodity, so Nvidia wins on efficiency forever" logic. Tokens, he argues, are not fungible: "a token from one model is not the same as a token from another model," because different models need wildly different numbers of reasoning tokens to reach the same answer. What's fungible is the intelligence built from tokens. The key line for anyone long the chip complex: "Right now, there is a price umbrella that is downstream of the lack of compute... Nvidia's customers like SpaceX AI can turn around and resell compute at high margins as well" (Tech Brew Ride Home, 2026-07-20). Translation: today's fat margins up and down the stack exist because demand exceeds supply. Thompson expects inference to grow much faster than training, frontier labs to eventually cut prices and "make it up in volume," and, critically for custom silicon, notes that "companies like Microsoft are increasingly obsessed with helping companies run their own models." That "run your own model" instinct is the same one driving every ASIC program.

5. The capex reckoning: Alphabet's negative free cash flow rewrote the narrative. [NEWS/PRINT] For two years, raising AI capex was a reason to buy these stocks. This week it became a reason to sell them. Alphabet raised full-year 2026 capex guidance to $195–205 billion, up from $180–190 billion, the CFO on the call: "The increase in the range is primarily due to an acceleration in the delivery of capacity to meet growing demand... we continue to expect our CapEx to increase significantly in 2027" (Squawk on the Street, 9am, 2026-07-23). Google spent ~$45 billion in the quarter alone, and free cash flow came in at negative $5.9 billion, the first negative FCF quarter in the company's two decades as a public company (The Rundown, 2026-07-23). Cramer's discomfort was audible: Alphabet is "a radically changed company... I don't want to be in a stock that owes that much money," noting no buybacks for a second straight quarter and that Google's 100-year sterling bond issued in February is already "down 7% since issuance" (Squawk on the Street, 9am).

Steve Eisman tied it together: adjusted EPS of $285 vs $231 a year ago but a miss versus the ~$289 expected, cloud up 82%, FCF negative, capex hiked, "the market is beginning to lose patience with all this crazy spending... the terms of debate on AI have truly shifted. A year ago, it was all rah-rah for AI. When companies raised their capex budgets, the market cheered." Not this week: Alphabet down ~7% Thursday, Tesla (which also went FCF-negative) down 14.5%, and the Nasdaq off more than 2% (The Real Eisman Playbook, 2026-07-24). Why a chip newsletter cares: if the market stops rewarding capex, hyperscalers get more cost-disciplined, and cost discipline is the structural argument for building your own cheaper silicon instead of buying merchant GPUs at 75% gross margin. The reckoning is bearish for the multiple and bullish for the ASIC thesis at the same time.

6. The financing-plumbing bear case got specific, and scary. [PUNDIT] On Big Technology (2026-07-24), the hosts unpacked Ed Zitron's argument about how the buildout is financed, and the numbers are worth writing down. Bloomberg estimates "over $500 billion in outstanding AI data center debt, at least $200 billion of it held by private credit, making up roughly 8% of outstanding private credit loans." Nikkei Asia reportedly puts the five big hyperscalers' (Meta, Google, Amazon, Microsoft, Oracle) accrued debt at "around $1.65 trillion... over the last 5 years, with an additional hundreds of billions of dollars worth of off-balance-sheet debt." The mechanism that rhymes with 2008: money routed through special-purpose vehicles doesn't count as capex, "Meta recorded $88.6 billion in capital expenditures, but that doesn't include the Hyperion SPV, which had its own $46 billion of exposure." And the punchline: Zitron estimates "70% of this capacity from Microsoft, Amazon, Google is OpenAI and Anthropic", two unprofitable companies as the "dual points of failure" underneath the whole thing (Big Technology, 2026-07-24). The line that captures the mood shift: "Right now, the burden of proof is on the skeptics, but once you have the slow trickle of disappointing information, then the burden starts to be on the optimists." This is the risk that would hit ASICs and GPUs alike, if the buyers can't fund the commitments, the whole chip debate becomes academic.

7. AMD, Anthropic, and Super Micro struck new deals, but the terms didn't surface. [NEWS/PRINT] Squawk on the Street flagged that AMD, Anthropic, and Super Micro "strike new deals" this week (Squawk on the Street, 10AM, 2026-07-22). One caveat: the episode's substantive discussion was about Chinese models (Kimi K3) and the Meta-vs-Alphabet debate, and I did not get the deal terms from the audio, so I'm flagging the headline, not fabricating numbers. Anthropic being the counterparty is notable, though: the same lab renting Google TPUs and buying compute across the ecosystem is now also linked to AMD. Anthropic is quietly becoming the single most important demand node in the entire accelerator market, watch it.

The debate

For the custom-silicon side (they take share and cap Nvidia's margins). This week the ASIC bulls got their first real evidence in a while, not just arguments. Google's TPU stopped being a purely internal tool and became an external revenue business, powering an 82% cloud growth rate and a half-trillion-dollar backlog. The next chip (Frozen) is funded and dated. And the macro turned in their favor: the moment the market stops applauding capex and starts demanding returns, the incentive to replace 75%-margin merchant GPUs with cheaper in-house silicon gets stronger, not weaker. Ben Thompson's "price umbrella" framing is the intellectual backbone here, Nvidia's margins reflect a compute shortage, and shortages don't last forever. When supply catches up, the merchant premium compresses and the captive-customer economics of a TPU or a Trainium look better and better.

For the merchant side (Nvidia keeps the pie). The counter is the treadmill, and it's still brutal. Vera Rubin's claimed 10x-per-watt improvement resets the bar Google's Frozen chip is chasing, and Frozen doesn't ship until 2028, against whatever Nvidia has by then. Frozen's whole design premise (freeze Gemini's architecture into silicon) is also a bet that could age badly if models keep changing shape. And the concentration problem cuts against the "TPU is a business now" story: a lot of that external demand is one customer, Anthropic. Meanwhile, Nvidia striking a $20B deal to license Groq's inference tech shows it intends to defend the exact inference beachhead the ASICs are storming.

The wildcard both sides still funnel to: Google, and now specifically the timeline. Last week the binary was "will Google sell TPUs externally?" This week the answer flipped to "yes, it started." So the debate advances to the next question: can Google's silicon roadmap (TPU today, Frozen in 2028) stay within striking distance of Nvidia's, or does the treadmill grind its edge to "50% better" the way it did to AMD? And hanging over all of it is the financing question, if the money underwriting the buildout gets thinner, both sides lose.

Where I come out this week: the scoreboard finally moved, and it moved toward the one ASIC that was always going to matter, Google's. The TPU is a real external business now, which is a genuine, dated fact rather than a pundit's hunch. But the same week handed the merchant camp its rebuttal (Vera Rubin's spec gap) and the whole complex its real risk (Alphabet's negative free cash flow and the SPV-debt plumbing). Net: I'm more constructive on Broadcom-as-arms-dealer and Google's TPU franchise than I was a week ago, more cautious on the multiple of everything downstream of the buildout, and increasingly convinced that the number to watch isn't a benchmark, it's free cash flow.

Stocks in play

  • GOOGL (Alphabet). [NEWS/PRINT, the week's main event] Bull: Cloud +82% to $24.8B (beat the 63% Street bar), $514B backlog, and the TPU is now generating external revenue from Anthropic and other labs, the exact monetization move that makes the custom-silicon thesis real (The Rundown; Limitless). The Frozen chip gives it a 2028 roadmap. Bear: first-ever negative FCF (-$5.9B), capex hiked to $195–205B with more coming in 2027, no buybacks, and ~two-thirds of backlog growth is a single customer (Anthropic) (Squawk 9am; Squawk 7/22). Watch: whether TPU external sales are disclosed as a line item, and next quarter's FCF.

  • NVDA (Nvidia). [PUNDIT] Bull: Vera Rubin's claimed 10x tokens/watt, 336B transistors on TSMC 3nm, and ~4x GB300 baseline reset the bar every ASIC must clear, plus the $20B Groq license shows it's defending inference (Limitless; Tech Brew). Bear: Ben Thompson's "price umbrella", margins ride a compute shortage that inference growth will eventually relieve; and if the capex reckoning slows hyperscaler buying, the merchant premium is first to compress (Tech Brew). Watch: an official Vera Rubin ship/availability confirmation to replace this week's pundit-sourced specs.

  • AVGO (Broadcom). [READ-THROUGH] Bull: if the custom-silicon field narrows to "Google's TPU wins," Broadcom is still the arms dealer building it and owns the Ethernet networking Google can't easily replace, and TPU external sales scaling means more XPU volume. Bear: no fresh Broadcom-specific commentary surfaced this week, and Baker's standing caveat ("Google can take the TPU away from Broadcom whenever they want") still applies. Watch: any disclosure on Broadcom's content per external-sold TPU.

  • AMD. [NEWS/PRINT] Bull: struck a new deal with Anthropic and Super Micro this week, a sign the necessary "second source" is winning real volume (Squawk 7/22). Bear: terms weren't disclosed on the podcast, and last week's cautionary tale stands, better node, better HBM, and it still trails Nvidia. Watch: the actual AMD-Anthropic deal terms.

  • META. [PUNDIT] Bull: analysts openly prefer Meta to Alphabet in the hyperscaler trade, its spend is "largely justified by what it does for their own business, not for services it sells to others," ad revenue growing ~30% on a $200B base, and "the ROIC pencils out very nicely" (Squawk 7/22). Bear: the SPV-debt discussion singled Meta out, $88.6B reported capex excludes a $46B Hyperion SPV exposure (Big Technology). Watch: Meta's capex number and any MTIA/Iris deployment color when it reports next week.

  • MSFT (Microsoft) / AMZN (Amazon). [PUNDIT] Bull: both report next week and the bar just got interesting, if either shows cloud acceleration like Google's, the capex gets forgiven. Bear: "the pressure is now on Microsoft, Amazon, and Meta on what they do with their CapEx" after Google refused to blink (The Rundown); and the "run your own model" instinct that Thompson flagged points straight at their Maia/Trainium programs. Watch: next week's prints, capex guides and any Maia 2 / Trainium 3 mentions.

  • INTC (Intel). [NEWS/PRINT, new to the board] Bull: a genuine positive surprise, EPS of $0.42 vs a $0.10 loss a year ago, data-center sales up 59%, best revenue growth in 15 years (The Real Eisman Playbook), and a real custom-silicon angle via advanced packaging (see Read-throughs). Bear: Eisman's caveat, a good Intel quarter "isn't going to quell AI nervousness," and the packaging optimism is priced in, so execution has to follow. Watch: whether the reported Intel-Google TPU packaging agreement turns into disclosed volume.

  • MRVL / ALAB / ARM / Alchip. [COVERAGE GAP, third week running] No fresh Marvell ASIC-roadmap, Astera Labs connectivity attach, Arm/Neoverse royalty, or Alchip color surfaced this week despite dedicated searches. This is now a persistent quiet stretch for the connectivity/IP layer, I expect it to re-surface as the 2027–28 designs (Frozen, TPU v7, Trainium 3, Maia 2) actually ramp and someone has to move the data between all those accelerators.

Read-throughs

  • TSMC / advanced packaging, still the door everything funnels through, but Intel just wedged a foot in. The most useful semis deep-dive of the week came from Chip Stock Investor, and it reframed packaging as a live competitive front. The core point: "Nvidia owns the training market... and it uses up most of TSMC's capacity in advanced packaging with the CoWoS technology" (Chip Stock Investor, 2026-07-21). That scarcity is the opening for Intel's alternative packaging (EMIB and Foveros), which is simpler and doesn't need a silicon interposer, and, crucially for our beat, the hosts noted Intel "has announced some agreements with, for example, Google and its TPUs possibly getting manufactured using processes like these simplified advanced packaging to get the HBM really close to those TPUs" (Chip Stock Investor). Read-through: CoWoS remains the choke point gating Nvidia and the ASICs alike, but inference-focused custom chips (TPU, Frozen) may not need CoWoS's full complexity, which is a modest structural positive for Intel Foundry and a reminder that packaging capacity, not just wafers, is where the bottleneck lives.

  • Inference demand looks structurally robust, which is the real bull case for volume across all silicon. On Forward Guidance, Steve Hou walked through his firm's compute indices and made a point that matters for every accelerator maker: even five-year-old A100 GPU rental rates "are going up and have stayed up, has not come down at all... that just tells you just how robust inference demand is" (Forward Guidance, 2026-07-22). His token-price index (an "expenditure-weighted price index... almost like the PCE for AI") has been mean-reverting as buyers shift from "token maxing" to "token efficiency" and smart model-routing, which coincided with the stock selloff. The read-through: demand for raw inference compute isn't the worry; the worry is whether cheaper models and smarter routing compress the price of intelligence enough to challenge how the capex is financed. That's a margin story, and it pressures the highest-margin link first (merchant GPUs).

  • HBM, still the toll booth under every accelerator. No fresh memory-maker prints on the beat this week, but the thread runs through everything: Google's Frozen and Intel's EMIB pitches are both partly about getting HBM physically closer to the compute die to cut data movement (Chip Stock Investor), and Vera Rubin's 22 TB/s memory bandwidth is the spec Nvidia leads on (Limitless). Whoever wins the logic war, HBM (Hynix/Micron/Samsung) gets paid on every unit.

  • The financing plumbing, the fault line under both merchant and custom silicon. Alphabet's negative FCF, the >$500B AI data-center debt stack, the SPV structures keeping ~$46B of Meta exposure off the capex line, and ~70% of hyperscaler capacity riding on OpenAI + Anthropic all point the same way: the dollars underwriting this buildout are getting thinner and more concentrated (Big Technology; The Real Eisman Playbook). This is the shared risk to every name in this newsletter. If the buyers blink, it won't matter whose chip is 10x more efficient.

What changed vs last week

Last week (issue dated July 20) was a debate week: the smartest voices, Gavin Baker, the a16z infrastructure expert, Jim Chanos, argued about custom silicon, and the consensus takeaway was "most ASICs will fail; the only real threat to Nvidia is Google's TPU, if Google decides to sell it externally." No new design win landed; the news was opinion. This week, three things concretely changed:

  • The TPU thesis went from "if" to "it's happening." Last week Baker's provocation was hypothetical, Google could sell TPUs on the open market. This week Alphabet's Q2 print and follow-on reporting showed Google started selling TPUs to external labs (Anthropic and others) in Q2, a step-change from Q1's internal-only use (Limitless). And a next-gen chip (Frozen v2, ~6–10x efficiency, 2028) got reported (Tech Brew Ride Home). The one ASIC the smart money said mattered is the one that just advanced.

  • The financing risk went from thesis to tape. Last week Chanos framed the danger as hyperscaler return-on-capital sliding toward 10%, a slow-moving worry. This week it arrived: Alphabet posted its first-ever negative free cash flow (-$5.9B), the market punished the stock ~7%, and the SPV-debt plumbing (>$500B of AI data-center debt, ~70% tied to OpenAI/Anthropic) got a full airing (The Real Eisman Playbook; Big Technology). Eisman's summary, "the terms of debate on AI have truly shifted", is exactly the mood change.

  • A new name joined the board: Intel. Last week Intel wasn't in the conversation. This week it printed its best revenue growth in 15 years, with data-center up 59%, and surfaced a real custom-silicon angle via advanced packaging and a reported Intel-Google TPU packaging agreement (Chip Stock Investor; The Real Eisman Playbook).

What did not change: still a coverage gap on the connectivity/IP layer (Marvell, Astera Labs, Arm, Alchip) for the third week running; still no published ASIC benchmark beating Blackwell in the wild; and still no hard external-shipment numbers for Trainium, Maia, or MTIA/Iris. The difference this week is that, unlike last week's pure debate, we got actual prints, and they leaned toward the custom-silicon thesis on substance while the market leaned bearish on valuation.