Newsletter · · Ashutosh Agarwal

AMD Ships the Helios Rack as Capex Judgment Day Nears - AI Accelerators - Week of July 23, 2026

AI Accelerators hardware newsletter for the week of July 23, 2026. Chip stocks bounced into the biggest earnings week of the year just as AMD finally shipped a full 72-chip rack called Helios and a multi-billion-dollar Anthropic deal, while the whole AI trade now hangs on one number: how much the hyperscalers say they will spend, with Bank of America modeling Alphabet at $195B this year and near $300B next.

AI Accelerators

Week of July 23, 2026: AMD Ships the Helios Rack as Capex Judgment Day Nears

GPUs, Custom Silicon & Optics: a twice-weekly read on the chips behind the AI build. Issue 011. Research window: July 16–23, 2026.

Monday's issue was about the chips breaking. This one is about the bounce, and the test that comes right after it.

After a brutal two weeks, the chip stocks steadied and then rallied hard into what is shaping up to be the single biggest earnings week of the year. But the rally only sharpened the real question, because almost every dollar of the AI trade now hangs on one number that nobody has actually seen yet: how much the giant cloud companies (the "hyperscalers," Amazon, Google, Microsoft, Meta) tell us they will spend on AI in the next twelve months. Alphabet reported after the closing bell on Tuesday, and the whole market treated it as the opening statement in a referendum on the entire build.

Underneath that macro drama, two genuinely new hardware stories broke this week. First, AMD finally has a product that answers Nvidia's best system head-on, a full rack of 72 chips called Helios, and, on the same day, a multi-billion-dollar deal to put its chips inside Anthropic (the maker of the Claude AI models). Second, a fresh look inside the memory-chip mania: one supplier has already pre-sold roughly 40% of next year's profit to a dozen-odd customers, even as its stock fell 20%.

Let's get into it.

TL;DR

  • AMD ships a rack, at last. AMD is launching Helios, its first "rack-scale" system, 72 chips wired to act as one brain, aimed squarely at Nvidia's top-end product. It is the first time AMD sells the whole rack as one finished thing instead of leaving customers to build it. Analyst take on The Information's TITV: genuinely competitive, priced a touch below Nvidia, but "you could argue" it is Nvidia's playbook a year or two late.
  • AMD + Anthropic. Per Wall Street Journal reporting relayed on CNBC, AMD is signing a server deal with Anthropic worth "tens of billions," taking a $5 billion stake in Anthropic, and may even backstop Anthropic's future data-center leases, the same circular "we'll invest in you so you'll buy our chips" pattern we've seen from Nvidia and OpenAI.
  • Capex is judgment day. Bank of America now models Alphabet spending $195B this year and nearly $300B next year, with the big four clouds together eating 98% of their operating cash flow. One analyst called it "heads I lose, tails you win": cut spending and you admit the returns are not there; raise it and investors panic about debt.
  • A heavyweight new bear. Interactive Brokers founder Thomas Peterffy: "much of the compute capacity that we are currently building will be a write-off in 3 to 5 years."
  • Memory's split screen. Prices for memory chips are up 300–500% in nine months and still climbing (+20% in July alone), yet the stocks fell about 20%. SK Hynix has reportedly pre-sold about 40% of its 2027 profit to just 13–15 customers. Samsung's Q1 operating profit was up 756%.
  • TSMC's tell. The world's most important chip factory raised its building budget above $60B and warned margins would dip, and its stock fell on record profits. The deeper story: making these AI chips needs three times the factory space of a phone chip.
  • Google's next chip has a name: "Frozen." Reportedly 6–10x more power-efficient than today's Google chips, arriving as soon as 2028, by literally baking part of the AI model into the silicon.

The developments that matter

1. AMD finally has a rack, and a customer to match

For six issues running, this newsletter has flagged the same hole: no real news from AMD, the only company with a plausible shot at loosening Nvidia's grip on AI chips. This week the news arrived, twice.

The product is Helios. On The Information's TITV, "US AI Policy Debate Reignited by Kimi K3... AMD's New Helios AI System" (July 21), Creative Strategies analyst Austin Lyons explained why it is a real step, not a press release. Until now, the biggest chunk of AMD you could buy was a server holding eight GPUs (the workhorse AI chips); if you wanted more, you strung servers together yourself. Helios changes that: it is "an entire rack full of 72 GPUs, all of it interconnected, all of them acting as one system," and, crucially, AMD sells it "as a whole product together for the first time."

Why does one-big-brain matter? Because the newest AI models are too large to run well on a handful of chips talking slowly to each other. Nvidia's answer is a rack called the NVL72 (72 chips acting as one); Helios is AMD's mirror of it. On price, Lyons expects "NVIDIA type prices, but just a little bit less," AMD's standard way of getting in the door. He also made AMD's honest counter-pitch: AMD sells a wide range of ordinary processors (its "Epic" CPU line, some with up to 256 cores), while Nvidia offers essentially one. For the new wave of "agentic" AI, where you run hundreds of software agents at once, that variety can genuinely matter.

But Lyons did not oversell it. Asked whether Helios is just "Nvidia's roadmap a year or two later," he answered: "that is fair pushback... you could argue that it is." And on the thing that would matter most, a customer standing up to vouch for it, he was blunt: "you'd love to see customers get up on stage and say we're excited to deploy this, so it's not just AMD talking its book."

Which brings us to the second piece. On Squawk on the Street, "10AM Hour: AMD, Anthropic & Super Micro Strike New Deals" (July 22), the hosts relayed Wall Street Journal reporting that AMD is signing a server deal with Anthropic "worth tens of billions of dollars," is taking a $5 billion stake in Anthropic, and "may also potentially backstop Anthropic's future data center leases." (Flag: this is WSJ reporting as described on-air, not an AMD announcement.) If it holds, it is a big deal for two reasons: it is a marquee AI lab publicly committing to AMD, and it is another example of the strange new financing loop where the chipmaker invests in the customer that then buys its chips, exactly what Nvidia did with OpenAI.

The same segment carried two more data points worth your book. Super Micro, which builds the servers these chips go into, said it took in over $60 billion in new orders in its fiscal fourth quarter, a record backlog. And SandboxAQ CEO Jack Hittery noted Anthropic separately signed with Elon Musk's SpaceX for what he described as "$45 billion spend over a 3-year period" (his characterization), calling compute "the lifeblood of these LLM companies."

Why it moves numbers: AMD's stock story has always been "someday it takes a real slice of the AI-chip market." Helios plus a named lab customer is the first hard evidence of "someday" turning into an order book. Watch the two-sided risk: real competition is good for AMD's revenue but bad for Nvidia's fat margins, and the circular financing makes both companies' "growth" partly self-funded.

But keep the asterisk: every word above came from an analyst or a reporter. AMD itself has not put an executive on a podcast to make the case, so the bull case for AMD still rests on secondhand accounts, not the operator's own voice.

2. Capex judgment day: the number the whole market is waiting on

Strip away the noise and the AI-chip trade now rests on one thing, whether the cloud giants confirm, and raise, their spending plans. This week that stopped being abstract.

On Squawk on the Street, "11AM Hour: Interactive Brokers Founder" (July 22), CNBC laid out the stakes ahead of Alphabet's report: Alphabet already more than doubled its building budget last quarter, and Bank of America now sees it spending $195 billion this year and nearly $300 billion next year. The kicker: across the big cloud companies, that spending is now eating 98% of their operating cash flow, the money the business actually generates. Alphabet has raised more than $140 billion in debt and stock since October, and "some models show free cash flow turning negative next year."

The Exchange, "John Paulson Stays Gold, Alphabet After the Bell" (July 22) captured the bind neatly. Analyst Rohit Kulkarni pegged the Alphabet "whisper number" at $200B this year and "at least $300B... could be as much as $400B" next year, with a growing worry that "the next $100B are not going to be as efficient as the previous $100B" because so much now goes to expensive memory. His line for the whole group: it is "heads I lose, tails you win." Trim spending and you are admitting the returns are not showing up; boost it and investors fret about debt and dilution. He thinks Google gets a pass, it has a valuable stake in SpaceX, recently raised $100B, and holds a stake in Anthropic that could go public, but "not a lot of other guys could."

Here is the plain-English version of why great earnings can still spook people, from Deutsche Bank strategist Binky Chadha on the same Squawk segment. S&P 500 profits are set to grow an all-time-record about 26% this quarter, but of that, about 11 percentage points come from semiconductors alone. As CNBC's Mike Santoli put it, roughly "$700 billion of what would have been free cash flow plus $400 billion raised in the debt market is going into somebody's earnings." Chadha's one-word reply: "Absolutely." Translation: the chipmakers' record profits are, in large part, the clouds' record spending. The bull case and the bear case are the same fact seen from two ends.

And the fresh heavyweight bear this week was a serious one. Interactive Brokers founder Thomas Peterffy (age 82), on the same Squawk 11AM segment: "much of the compute capacity that we are currently building will be a write-off in 3 to 5 years. The big AI companies are faced with the choice of either continuing to build capacity and incur likely losses or drop out of the race altogether. The resulting overcapacity will cause severe compression in margins. Infrastructure developers and the old clouds will have trouble refinancing. Stock prices will come down." He is bearish "for the next 5 years or so," but added that after a shakeout, "10 years down the road, it will be terrific."

Why it matters: this is no longer a philosophical bubble debate. It is a scheduled, dated event, earnings guidance, one company at a time, and the chip stocks will trade tick-for-tick on whether the clouds keep the spending taps open.

3. Memory: the split screen between the price and the stock

The memory-chip names (the companies that make the DRAM and high-bandwidth memory that sits next to every AI chip) gave us the week's strangest picture: the product keeps getting more expensive while the stocks fall.

On Limitless: An AI Podcast, "Elon Will Handle the Energy Bottleneck Himself" (July 22), the hosts laid out the disconnect with real numbers. Memory prices are up "on average three to 500% over the last nine months," and just in July "the average price of DRAM is up almost 20%," while the stocks fell about 20% over the same stretch. Their punchline on demand: SK Hynix has signed long-term contracts where "between 13 to 15 customers secured around 40% of their projected profit for next year," meaning it has effectively "sold out their supply for the entirety of 2027." They flagged Micron trading at "7x forward earnings... and 85% gross margins," and argued the real bottleneck, building the factories, "isn't going to get unlocked until around 2030." Their read on the sell-off: over-leveraged retail traders in Korea unwinding (the market shed roughly $1.5 trillion), not any change in demand.

That the earnings are real is not in dispute. Patrick Boyle On Finance, "The World's Best Stock Market Is Also Crashing!" (July 19) put hard figures on it: Samsung's Q1 2026 operating profit rose 756% to 57.2 trillion won; SK Hynix's revenue rose 198% and operating profit 405%. Boyle's memorable frame for the current shift to usage-based AI pricing: it is like "installing a water meter in a house that used to get its water for free, people use less." He also connected the boom to the whole Korean market's wild ride (the KOSPI index roughly doubled to a June peak, then fell 27%), noting the entire thing has become "a leveraged bet on the server budgets of about 4 American corporations."

The steel-man on the other side came from Merryn Talks Money, "Is Asia's AI Boom Creating the Next Investment Bubble?" (July 20): a fund manager warned that Samsung's earnings grew 19-fold year-over-year and shares still fell (expectations were higher), and that the memory stocks price in "years of constrained supply," a setup that breaks badly if hyperscaler demand cools or supply arrives faster than expected, especially with Chinese rivals undercutting on cost.

The read-through you can use came from The Catalyst by Softchoice, "The AI PC Episode" (July 16): data centers are expected to consume 70% of the world's memory chips in 2026, up from 20–30% before. That is why memory prices jumped 80–90% in a single quarter (per Counterpoint), and why the same podcast expects everyday PC and phone prices to rise 10–20% by the end of 2026. The AI build is quietly reaching into the price of your next laptop.

4. TSMC's quiet tell, and why AI needs three times the factory

TSMC, the Taiwanese company that actually manufactures Nvidia's and most everyone's advanced chips, reported blowout numbers and its stock fell. The Circuit, "Ep 184: Earnings TSMC/ASML/AEHR" (July 20) explained the tell. TSMC beat, guided up, and raised its building budget into "the 60s" (above $60 billion), with 70–80% going to the most advanced manufacturing, including the "packaging" step (called CoWoS) that stitches an AI chip and its memory together. It also warned margins would come under pressure. Records plus rising costs plus a margin warning equals a stock that falls.

The bigger insight from hosts Ben Bajarin and Jay Goldberg is one worth holding onto: an AI chip is physically enormous compared with a phone chip, so the same demand requires "four to five additional foundries" (chip factories). The machines that make the newest chips are "at least two and a half times the size of prior nodes," so, in Bajarin's words, "you got to build three more buildings to make less chips than you did in a smartphone." That is the structural reason the capex numbers keep climbing, and why the factory-equipment makers (Applied Materials, KLA, Lam Research) are, as Goldberg put it, growing "at exponentials" even as their stocks fell for the week, "nobody cares."

The Circuit also gave the bulls a rebuttal to the "it'll all be a write-off" bears (see Peterffy above). Bajarin's back-of-envelope model: at about $6 an hour of rental and 75% usage, a cloud company pays back the cost of its Nvidia chips "in a little over two years," and pays back an entire billion-watt data center, power, building and all, "in 7.1 years." The catch he is honest about: that only holds "as long as demand keeps going up."

5. Custom silicon: Google's "Frozen" chip, and the Cerebras land-grab

The most interesting custom-chip news came via Tech Brew Ride Home, "It's All China AI All The Time" (July 20), reading from reporting by The Information. Google is building a new chip nicknamed "Frozen," which its own engineers project could be "10 to 6 times more efficient" than Google's current homegrown chips (measured in useful AI output per unit of power), arriving "as soon as 2028." The trick behind the name: it "permanently etches parts of the model into the silicon," hard-wiring Google's Gemini AI so the chip does less work per answer. It is a bet that Google will not radically change how Gemini is built. The same segment noted OpenAI, Microsoft, and startups SambaNova and d-Matrix are all chasing similar efficiency-focused chips, and, a striking line to flag, that "Nvidia itself struck a $20 billion deal in December to license technology from Groq," a startup that designs inference chips (reported by The Information as relayed on the podcast; not independently confirmed here).

On the challenger front, The 7investing Podcast, "Cerebras vs. NVIDIA: The $30 Billion AI Inference Battle" (July 20) put numbers to Cerebras's rise: an OpenAI order of roughly $10 billion through 2028 for its giant wafer-scale chips (used for fast, "latency-sensitive" AI responses), plus a potential about $20 billion Amazon deal for AWS (both reported and discussed by the analysts, not confirmed by the companies). That dovetails with the DigiPower X operator interview we flagged Monday: on Wall Street Unplugged, "How DigiPower X's $1B Cerebras deal is powering execution" (July 17), CEO Michel Amar detailed his $1.1B (expandable to $2.5B) co-location contract with Cerebras, with the first phase due in Alabama in December 2026.

The big debate, steel-manned

The bubble argument matured this week from name-calling into two coherent, testable views. Both can be held by serious people.

The bear case: the returns are not there, and the gear is short-lived. Beyond Peterffy's "write-off in 3 to 5 years," the bears got specific. On Excess Returns, "It Only Happens at Bottoms" (July 18), derivatives veteran Andy Constan argued the whole boom funnels into a handful of firms, "three or so companies that sell memory," "three or four that sell compute chips," "one company that makes raw silicon," plus a "Fab Five" of equipment makers, and there simply is not enough economic output (GDP) to justify what all that invested capital is supposed to earn back. On Thoughtful Money (July 18), Lance Roberts tallied the week's damage, Applied Materials −12%, Lam −12%, Micron −13%, Intel −16%, Western Digital −21%, and cited veteran skeptic Fred Hickey's claim that, once you account for the coming wave of gear write-downs, the group's real price-to-earnings ratio is nearer 67 than the headline number. And Jim Chanos, on RiskReversal Pod, "The Math Ain't Mathing" (July 17), repeated his core worry: the return on each new dollar invested is sliding from 40% toward 20% and, he thinks, on to 10%.

The bull case: this is scarcity, and it is defensive, so it lasts. On The Exchange (July 22), Morgan Stanley's Andrew Slimmon gave the cleanest version: "invest in scarcity. There's a scarcity of memory chips. There is a scarcity of compute power... the big hyperscalers, they're not hyped at all." His answer to the "commodities always crash" retort: it comes down to how long until supply catches demand, and "the dot-com [boom] blew when everyone that needed a router bought a router... I just think we're not there yet." On Monetary Matters, "...Not AI | Russell Clark" (July 22), Russell Clark added a structural twist: the spending is not a bubble, it is defensive, Google, Microsoft, Amazon and Meta protecting their existing profit engines against the threat of Elon Musk's compute, the way carmakers had to answer Tesla. In that view the capex simply does not stop, and Nvidia's pricing stays firm "like 1970s oil." And Morgan Stanley's economics team, on Thoughts on the Market (July 21), framed it as a multi-year engine: US cloud AI spending rising to $1.2–1.4 trillion through 2028, with about 60% of it flowing to Asian suppliers, a cycle they see running "three to four more years."

The swing factor is exactly what makes this week different: it is no longer a debate you can argue forever. It resolves, in installments, at each earnings report. As Bloomberg Intelligence's Mandeep Singh put it on Bloomberg Intelligence, "Chip Stocks Sink Into Bear Market Territory" (July 17), the sell-off hit "despite unchanged fundamentals," and he expects the capex numbers to actually increase when companies report. If he is right, the bulls win the next round. If Peterffy is right, the write-offs start showing up in 3–5 years, not next week, which is precisely why this can run a long time before it is settled.

Names in play this week

  • AMD: Helios (72-chip rack) plus the reported $5B Anthropic stake and lease backstop. The clearest "AMD is finally competing" evidence in months.
  • Nvidia: the reference point for everything above; notably, on The Compound and Friends (July 17), one host flagged Nvidia's price-to-earnings ratio has "compressed to 18x, below the S&P average" with no slowdown in profits, the market simply "bored." Also the reported $20B Groq licensing deal (flagged, unconfirmed).
  • Anthropic: the customer at the center of the week (AMD, plus a reported $45B SpaceX compute deal).
  • Super Micro: a record over $60B in new orders; the clearest single read on server demand.
  • SK Hynix / Micron / Samsung: prices up 300–500% over nine months; SK Hynix about 40% of 2027 profit pre-sold; Samsung Q1 operating profit +756%; all three stocks down about 20–30% anyway.
  • TSMC: building budget raised above $60B, margin warning, stock down on records.
  • Applied Materials / KLA / Lam Research: the factory-equipment names quietly compounding while ignored.
  • Google (Alphabet): the "Frozen" custom chip (2028), and the capex report that set the tone for the whole season.
  • Cerebras: reported about $10B OpenAI and about $20B Amazon inference deals; the DigiPower X buildout is the physical footprint.

Read-throughs

  • From the chip trade to your laptop: with data centers set to soak up 70% of the world's memory in 2026, everyday device prices are expected to rise 10–20% by year-end. The memory shortage is no longer just a stock story.
  • From capex to the whole tape: because about 11 of the S&P 500's record about 26% profit growth comes from semiconductors, and that growth is literally the clouds' spending, a single soft capex guide from a hyperscaler can whip the entire index, not just the chip names.
  • From Korea to everything: as Patrick Boyle put it, two Korean memory firms have become "a leveraged bet on the server budgets of about 4 American corporations." When those four report, Seoul feels it.
  • The circular-financing web tightens: Nvidia to OpenAI, AMD to OpenAI, and now AMD to Anthropic (with a $5B stake and possible lease backstops), plus Anthropic to SpaceX. More of the demand propping up chip revenue is being funded by the chipmakers and their partners themselves. That is great on the way up and reflexive on the way down, worth watching as a systemic risk, not just a growth story.

What changed since Monday

  • The frame flipped from break to test. Monday led with the chip index falling into "bear-market" territory and TSMC's blowout print. This issue leads with the bounce into the biggest earnings week, and the realization that the entire trade now hinges on hyperscaler capex guidance, starting with Alphabet after the bell.
  • AMD's news finally broke. After six issues of nothing from AMD, we finally got a real product (Helios) and a marquee customer (Anthropic, with a $5B stake), though it all came from analysts and reporters rather than the company itself.
  • A new heavyweight bear (Peterffy) and a new structural bull (Russell Clark). Monday's referendum leaned on Chanos, Zitron, Boockvar. This issue adds Interactive Brokers' Peterffy ("write-off in 3 to 5 years") on the bear side and Clark's "defensive-capex, like 1970s oil" thesis on the bull side, plus hard fresh capex figures (BofA's $195B/$300B for Alphabet; 98% of operating cash flow).
  • Memory got concrete. Monday cited memory-share percentages; this issue has the mechanics, SK Hynix pre-selling about 40% of 2027 profit to 13–15 customers, prices +20% in July, the fab bottleneck lasting to 2030, plus the downstream device-cost shock.
  • New custom-silicon detail: Google's "Frozen" chip (a name and a 2028 date) and firmer Cerebras deal figures, neither in Monday's issue.