Newsletter · · Ashutosh Agarwal

AMD Beats Estimates and the Stock Still Falls - AI Accelerators: GPUs, Custom Silicon & Optics - Week of August 6, 2026

A synthesis of what hardware and semiconductor podcasts said from July 30 to August 6, 2026, as AMD beat estimates and still fell as much as 10%, the market began separating AI winners from losers, and an operator called memory pricing a 100-year flood, for the week of August 6, 2026.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of August 6, 2026: AMD Beats Estimates and the Stock Still Falls

A week ago the story on these podcasts was fear. The AI trade looked broken: Alphabet had just gone cash-negative for the first time since it went public, a young hedge-fund star had blown up betting on exactly this build-out, and chip stocks were in free fall. Then the rest of Big Tech reported, and the mood flipped again, hard. By Tuesday, one host put a number on it: the Nasdaq had "jumped 9.3% in just four trading days, adding roughly $3.5 trillion in market value" (The Rundown, Aug 5).

Here is the thing worth holding onto. The market did not decide "AI is fine" or "AI is a bubble." It started doing something more grown-up: it began separating the companies that can show a return from the ones that can't. Microsoft and Amazon proved their spending is turning into cash and got rewarded. Meta spent just as freely, couldn't explain the payoff, and got punished. And the marquee hardware event of the week, AMD's earnings, showed the same reflex in miniature: AMD beat on almost every number, landed every major AI lab as a customer, and still fell as much as 10%, because when a stock is priced for perfection, "very good" is a disappointment.

That's the through-line of this issue. Let's dig in.

(A quick note on words we'll use. "Capex" is capital expenditure, the money these companies spend building data centers and buying chips. "HBM" is high-bandwidth memory, the ultra-fast, ultra-expensive memory stacked next to an AI chip; it's roughly half the cost of a modern GPU. A "hyperscaler" is a giant cloud operator, Microsoft, Amazon, Google, Meta. A "gigawatt" is a unit of electrical power; one gigawatt is roughly the draw of a mid-sized city, and it's become the new yardstick for AI build-outs. We'll define the rest as we go.)

TL;DR

  • AMD is the week's big hardware print, and the cleanest example of "not good enough." Revenue up 50% to $11.5B, data-center sales more than doubled to $6.7B, guidance above Street. Stock fell 6–10%. It trades near 50x forward earnings versus Nvidia's ~20x while holding ~4.5% of the data-center GPU market to Nvidia's ~95%. The story that matters: AMD's new "Helios" rack landed all five frontier AI labs (Microsoft, OpenAI, Meta, Oracle, Anthropic), the first time it's done that, but the real money is a Q4-into-2027 story, and one analyst says the design is "almost a year behind" Nvidia.

  • The market is now sorting winners from spenders. Microsoft (+8–15% on the day) and Amazon proved the spending pays off; Meta (−8 to −11%) couldn't. Same capex, opposite reactions.

  • Amazon's CEO gave the single most bullish operator line of the week: $220B of capex this year, pushed up by memory costs, and even that won't be enough to meet demand, in 2026, 2027, and looking into 2028.

  • Memory is a "100-year flood," per Apple's Tim Cook. SK Hynix reported more profit than revenue and is sold out of HBM through 2027, yet its stock had crashed. The gap between the fundamentals and the share price is the cleanest sentiment-vs-reality picture in the market.

  • The bubble debate got two heavyweight new voices: legendary short-seller Jim Chanos (bear: it's an accounting mirage) and investor Gavin Baker (bull: GPU rental prices are rising, so the spenders are actually under-earning).

  • New money is flying: Anthropic is finally building its own chip (with Samsung) and signed a $10B compute deal in Norway on Nvidia's next-gen chips; OpenAI is planning a >$30B Georgia site and a possible $500B, 10-gigawatt Ohio hub.

  • Networking woke up: Arista blew out earnings (+11–12%). Pure optics names stayed silent for another week.

1) The lead: AMD beats, lands everyone, and still gets sold

AMD reported after the bell on August 4, and on paper it was excellent. Revenue rose 50% year-over-year to $11.54 billion, ahead of the ~$11.31B Wall Street wanted. Adjusted earnings were a record $1.66 a share. The engine was the data-center business, where revenue "more than doubled to a record $6.7 billion" and now makes up 58% of the whole company, up from 42% a year earlier (The Rundown, Aug 5). Guidance for the current quarter came in at a midpoint of $13 billion, above the ~$12.5B consensus. CEO Lisa Su summed it up: "we delivered an excellent quarter. We had record revenue and profitability as data center revenue more than doubled" (Schwab Network, "EARNINGS ALERT: BKNG, SPCX, AMD," Aug 4).

The stock fell 6–10%.

Why? The most useful explanation came from a plain-spoken breakdown on the Elon Musk Podcast (Aug 5): "the stock had already doubled prior to this report, so… the market had basically priced in absolute perfection before the numbers even hit the wire." Put numbers on it, AMD trades near 50 times forward earnings, while Nvidia, its far larger rival, trades near 20 times. And AMD holds roughly 4.5% of the data-center GPU market against Nvidia's ~95%. When you pay that premium, the hosts noted, "merely beating current revenue estimates by, say, $500 million just isn't enough." Bernstein's Stacy Rasgon, who rates the stock a buy with a $600 target, agreed the quarter "looks fine… there's nothing wrong with it," and pinned the drop on "heightened expectations" (Closing Bell, Aug 4).

The genuinely new and important part is Helios. Helios is AMD's first "rack-scale" system, instead of selling loose chips, AMD now sells a whole refrigerator-sized rack that bundles its GPUs, CPUs, networking, and liquid cooling into one unit designed to go head-to-head with Nvidia's rack systems. And AMD announced that Microsoft, OpenAI, Meta, Oracle, and Anthropic have all committed to deploying it "at a gigawatt scale." As one show put it, "this is the first time that AMD has landed all five major frontier AI labs on a single rack platform. They have everyone signed up for it" (Practical News, Aug 5). The Anthropic deal alone is for up to 2 gigawatts of AMD GPUs, the same scale as Nvidia's largest training cluster. Lisa Su also told AMD's product event the total market for AI accelerators will reach $1.3–1.4 trillion by 2030, "approaching the size of today's entire semiconductor industry."

For readers who've followed this newsletter: this is the operator we've been waiting for. For eight straight issues, AMD's own executives were absent from these podcasts even as the company made headline-grabbing deals. This week AMD is finally front and center, with hard numbers and its CEO's own words on the record, even if the market's reaction was a shrug.

Now the two catches, because they're what a book actually needs to underwrite.

Catch one: the money is mostly not here yet. Rasgon: "the big AI ramp with Helios really is a Q4 story," with the follow-on chip (MI450) ramping "into next year." A sharper version came from Ben Palladino of BEP Research on The Information's TITV (Aug 5), who argued Helios's real high-volume ramp is "Q1 of 2027, which in AI data center land is… almost like a generation behind." His reason is worth understanding: Helios uses big, thick internal cables and a "double-wide" design, exactly the kind of physically awkward rack that gave Nvidia's own Blackwell system assembly headaches a couple of years ago. Nvidia's next design (Vera Rubin) moves away from those cables. So, Palladino argues, "you could effectively say that AMD Helios is almost like a year behind in technology knowledge." Time-to-market is everything here: customers plan a data center's architecture a year or more ahead, so a slow ramp means AMD misses the window.

Catch two: a real customer just said no. On SpaceX's first-ever earnings call, held the same evening, Elon Musk said SpaceX will build its AI infrastructure "exclusively with NVIDIA's Blackwell chips", a reversal from earlier comments that it would buy from both. As The Rundown put it, "that has to be a bit of a gut punch to AMD."

And notice one number that didn't move: AMD guided gross margin flat at 56%, even as its richer data-center mix should have pushed it up. Two culprits, both structural and both good context for the whole sector. First, HBM memory costs are climbing and "eat directly into the margin." Second, selling a whole rack means "you are selling the sheet metal, the heavy copper bus bars… the internal network switches, and the liquid cooling manifolds. None of those things carry semiconductor-level margins" (Elon Musk Podcast, Aug 5). The upside of racks is stickiness, once a customer designs a data hall around your rack's power and plumbing, ripping it out later is "construction work, not just IT work." AMD is trading a little margin for a lot of lock-in.

Why it matters for the thesis: AMD is now a credible "second source" that every major lab wants, a real change from Nvidia's near-monopoly. But the market is telling you the "second source" trade is priced for a rapid share grab that hasn't happened yet, and the physical ramp is a 2027 event. Palladino's blunt framing: "AMD is sort of squeezed in the middle," with Nvidia dominant above it and hyperscalers' own custom chips plus a swarm of startups nibbling below.

2) The market stops clapping for everyone and starts picking

The clearest way to understand the week is to line up three earnings reactions.

Microsoft: rewarded. Revenue up 18%, profit up 32%, and its Azure cloud business accelerated to 43% growth, faster than the prior quarter and faster than analysts expected, crossing $100 billion in annual revenue for the first time. Crucially, Microsoft is turning its spending into a business: 30 million people now pay for its Copilot AI assistant, up from 20 million a quarter earlier (Morning Brew Daily, Jul 30). It spent $41 billion on capex in the quarter (up ~70% year-over-year) toward roughly $190 billion for the year, yet still generated $20 billion of positive free cash flow and guided to stay positive next year.

Gil Luria, head of technology research at DA Davidson, called it "a narrative-breaking result" on Prof G Markets (Jul 30). His most useful point was about a single phrase from Microsoft's CFO, that capex would be "up year over year." Luria: "If she said down year over year, all hell would break loose in the market… all semi stocks would get cut in half." Instead, "she went right down the fairway." The message investors heard: Microsoft will keep spending because the returns are there, not out of fear. Luria's verdict: "Microsoft is the adult in the room… they're not doing it in a big game theory, game of chicken competition."

Meta: punished. Revenue actually grew a strong 28% to $60.8 billion, nearly twice Google's pace, and daily users hit an all-time high of 3.6 billion. But earnings missed badly, $6.18 a share against $7.20 expected, costs jumped 55%, and free cash flow collapsed to $784 million from $8.5 billion a year earlier (Schwab Network, Jul 30). The stock fell 8–11%.

Two honest readings, and it's worth holding both. Luria's bearish take: the report "was barely passable," and Mark Zuckerberg "had a whole hour to explain how he's going to monetize the massive AI investments. And he didn't really give us a firm answer." The gentler take came from investor Joseph Carlson, who's long the stock and bought more on the drop. He pointed out the earnings miss was mostly one-time items: a "$2.4 billion charge related to legal proceedings," "$1.18 billion of severance" from a May layoff, and faster depreciation of AI equipment. "When these three factors are taken out, their earnings per share were in line" (The Joseph Carlson Show, Jul 31). His broader thesis is that Meta is simply last in line for the same re-rating the market already gave Google, Microsoft, and Amazon, because, unlike the other three, Meta has no public cloud business to directly sell its AI capacity into.

Amazon: rewarded, and the loudest bull signal of the week. AWS grew 37% to a $169 billion annualized run rate, its fastest growth in 18 quarters. But the line that moved people came straight from CEO Andy Jassy, relayed on Squawk on the Street (Jul 31): "We now believe we will spend approximately $220 billion in cash CapEx in 2026, the higher cost of memory pushing this number up from our prior estimate of about $200 billion. But even at that amount, we will still not have enough capacity to meet all the demand we have in 2026. And I believe this dynamic will also be true in 2027." Jim Cramer's reaction captured why it landed: "Andy Jassy is the first of these execs who have said, look, this is where the money is." (Free cash flow was still negative $7.6 billion, the enormity of the build-out means even Amazon is now burning cash.)

Why it matters: The blanket "AI is a bubble / AI is a boom" question is being replaced by a company-by-company one, can you show the return? That's healthier for the durable hardware winners (whoever the cloud giants keep buying from) and more dangerous for anyone whose story is momentum. As Chip Stock Investor tallied it up, capex across the big spenders (plus Tesla and SpaceX) is running near $580 billion over the trailing year, heading toward ~$900 billion for 2026 and "possibly… $1.5 trillion in 2027" (Chip Stock Investor, Aug 3). The taps are not turning off. They are just being pointed at whoever can prove they'll fill the bucket.

3) Memory: an operator calls it a "100-year flood"

If you want the single most vivid line of the week, it came from Apple CEO Tim Cook, on what he said was his final earnings call, relayed by Jim Cramer: the memory situation is "a 100-year flood." Cramer's gloss made the stakes clear, Apple simply can't get enough memory chips, because hyperscalers are outbidding everyone. "It's not that the chips are there and they're not willing to pay. There are no chips for them." And nobody is promising 2027 gets better (Squawk on the Street, Jul 31). Remember Jassy's line from the section above: his capex went up specifically "because of the higher cost of memory." When the world's best supply-chain operator (Cook) and the world's largest cloud buyer (Jassy) both point at memory in the same week, that's a signal.

Now the paradox. The best deep-dive was on Limitless, "The Memory Selloff: Record Profits, Collapsing Stocks" (Jul 30). SK Hynix, the leading maker of the HBM memory that sits next to every AI chip, reported one of the great quarters in corporate history: revenue up 354%, and more net profit than revenue (net margins of 118%, thanks to gains beyond the core business). Analyst Ejaz: "They made more money in this quarter than they did in the entirety of 2025." And they are "sold out" of HBM not just for 2026 but into 2027.

So why did the stock crash (the Korean market fell more than 30%)? Three reasons, and they're a masterclass in sentiment versus fundamentals. First, SK Hynix technically "missed" one analyst's revenue target by ~$1.7 billion, because it's physically sold out and literally can't ship more. Second, being sold out through 2027 "eliminates a lot of the upside surprises," leaving only downside if a deal falls through. Third, a leveraged Korean ETF launched in mid-July turned an ordinary pullback into a cascade of forced selling; JPMorgan estimated that liquidation was "about 90% complete," implying a bottom is near. The market later agreed, Micron reportedly popped 20% as the memory names recovered into the week.

Why it matters: The screen (bear market in memory stocks) and the ground (sold out through 2027) told opposite stories, and the ground looks right. Since HBM is roughly half the cost of a GPU, a genuine, multi-year memory shortage is both a hard ceiling on how fast the whole AI build-out can go and the most direct way to play it. As Ejaz framed the whole cycle: the cash is flowing "from the hyperscalers to the semiconductor companies," and "as boring as the answer is," that still points at memory and chips.

4) The debate, steel-manned: Chanos vs. Baker

The bubble argument didn't fade this week, it got sharper, because two of the most respected voices on each side laid out full cases. The backdrop was a real-world casualty: the AI-focused hedge fund Situational Awareness, run by 25-year-old former OpenAI researcher Leopold Aschenbrenner, blew up spectacularly (reported at a ~$20B fund with losses variously put in the tens of billions, figures vary across shows and should be treated as reported, not confirmed). Cramer even blamed the "Leo effect" for part of SK Hynix's swing. It was the perfect prompt for both sides.

The bear: Jim Chanos, the short-seller who called Enron, on Prof G Markets (Jul 31). His case isn't "AI is fake"; it's "this is a capex boom, and capex booms end badly." His most important and least-understood point is an accounting one. The companies spending on data centers get to spread that cost over 5–10 years (depreciation), but the companies receiving the money, Nvidia, equipment makers, utilities, book it as revenue and profit immediately. That mismatch, he argues, is why S&P 500 profits are "growing somewhere like 28 or 29" percent when the economy would normally support 8 or 9. Strip out the mismatch and the earnings picture is far less glorious. His killer metric: the return on each incremental dollar the hyperscalers invest "has been pretty much cut in half over the past year and a half," with some now "earning in the teens." If that keeps sliding, "the C-suites in Silicon Valley… are going to be having some interesting conversations." His read on where we are: judging by the surge in new stock issuance (think SpaceX's IPO), "we're closer to a 99 type moment than a 97", i.e., near the top, not the start.

The bull: Gavin Baker of Atreides, on Invest Like the Best (Aug 4). His argument rests on a single ground-level fact that directly answers the bear: the price to rent a GPU is going up, not down. He described a startup that rented several thousand Blackwell chips at roughly "mid-$2 per GPU hour" and, seven months later, expects to pay "just under $4" for the identical cluster, a 50–60% increase. An inference-cloud company said on a podcast it plans to "pay 100% more for Blackwells when our contract expires." If used chips are re-pricing higher, then, in Baker's words, "all the hyperscalers are under earning", the opposite of a glut. He also dismantled the "cheap open-source models will kill demand" fear with a memorable line: "A token is a token." Whether a request runs on a frontier model or an open-source one, "you need the exact same amount of compute", open source just moves profit from the model makers into the infrastructure layer; it doesn't reduce the need for chips. (And those "free" models aren't cheap to run: he pegged Kimi K3 at "$1.1 to $1.2 million" to operate at scale.) On the scariest bear point, credit stress, with Meta's latest bond pricing wide and Nvidia's credit-default swaps "blowing out", Baker's counter is that most of the build-out is still funded from operating cash flow, and the widening is "just banks hedging their commitments."

Where they actually agree, and where you should watch, is the same place: the return on invested capital. Chanos says it's already halving; Baker says rising rental prices mean it's about to inflect up as old contracts re-price. That single variable, are the returns on this spending getting better or worse?, is the real debate. Everything else is noise around it.

One structural point from Baker worth filing away, because it changes the game versus past cycles: breaking a long-term supply agreement now carries a severe penalty. In the old world, a giant buyer like Apple could walk away from a memory contract with no consequences. Today, with four-plus AI labs competing for scarce chips, "you might blow up your entire business and your franchise by breaking" an agreement, because your supplier will simply hand your allocation to a rival. That fear of losing your place in line is, itself, a force keeping the spending, and the shortage, going.

5) Names and money in play

Anthropic finally builds its own chip. The company is standing up a custom-silicon team to design chips for its Claude models and is "scouting Samsung as a manufacturing partner," which "ends Anthropic's status as basically the last major AI lab without a public silicon program" (Practical News, Aug 5). For context, OpenAI shipped its Broadcom-designed "Jalapeno" inference chip in June, Google runs its own TPUs, and Meta is building its MTIA chips. Building custom silicon "takes 18 to 24 months and costs hundreds of millions of dollars per chip version", a slow, expensive margin play, not a quick fix, given Anthropic reportedly spends around $1.25 billion a month on compute.

The deal web keeps spinning. Anthropic also signed a $10 billion, six-year compute deal with Volta Infra, a Nvidia-backed startup founded in January by ex-Brookfield executives, for a 133-megawatt site in Norway stocked with Nvidia's newest "Vera Rubin" chips, at roughly $12.5 million per megawatt per year. Volta promptly raised ~$300 million (from A16Z and Altimeter) at a $2.4 billion valuation off the back of that contract, and is converting former Bitcoin-mining sites to AI (Tech Brew Ride Home, "OpenAI Has Receipts," Aug 4). Meanwhile OpenAI is planning a Georgia data center that "may cost more than $30 billion," and Nvidia is in talks to help it lease a $500 billion, 10-gigawatt hub in Ohio overseen by SoftBank. These are exactly the "increasingly interconnected web of dependencies" that the bears (and Bloomberg's own reporting, cited on the show) flag as circular, money and chips looping between the same handful of players.

Amazon's own chips are quietly a $25B business. AWS chief Matt Garman said the company's AI chip line has a $25 billion revenue run rate, and that its Trainium chips are "largely sold out through the end of next year," saving customers "20%, 30% off their inference costs" versus Nvidia on tuned workloads (Bloomberg Talks, Aug 3). It's the strongest evidence yet that custom hyperscaler silicon is a real, sold-out business, not a science project.

Challengers and pricing. Nvidia-rival startup Etched raised $300 million, and Jensen Huang published an open letter backing open-source AI (20VC, Jul 30). And AMD is reportedly raising GPU prices by at least 10% in August, following Nvidia, a small but telling sign of who has pricing power right now.

6) Networking wakes up; optics stays asleep

For weeks this newsletter has flagged the silence of the optical-networking names. This week one corner of networking spoke loudly. Arista Networks, the leading maker of the high-speed switches that wire data centers together, reported its first-ever quarter above $3 billion in revenue ($3.04B versus $2.83B expected), earnings of $1.02 versus $0.88, operating margins near 50%, and guidance well ahead of the Street. The stock jumped 11–12% (Schwab Network, Aug 4; Closing Bell, Aug 4). One analyst called it "Palantir-like… every single segment beat." Arista is the clearest listed read-through on hyperscaler networking spend outside Nvidia itself, and the one constraint flagged, demand outpacing supply, is the good kind of problem.

The negative space still matters, though. The pure optics names, Coherent, Lumentum, Astera Labs, Credo, were essentially absent from podcasts again this week. That's now a multi-week silence in a part of the supply chain (optical transceivers and co-packaged optics) that should be a direct beneficiary of every gigawatt built. Silence isn't bearish on its own, but it means the podcast tape is giving you no fresh operator signal there, worth noting rather than papering over.

7) Read-throughs

  • Power is still the binding constraint, and it's turning political. Morgan Stanley's team pegged a potential 38-gigawatt power shortfall through 2028, and noted local moratoria are mostly "a pause, not an outright ban", so the risk is to the pace of the build-out, pushing data centers toward rural areas, Canada, and Australia (Thoughts on the Market, Aug 4). In Texas, developers have proposed roughly 900 gigawatts of projects against a grid with about 90 gigawatts of total capacity, a tenfold oversubscription that pushed the governor to order audits (effectively ~12-month delays) (Telltales, Aug 5). The mid-Atlantic grid operator PJM's July auction cleared right at its price cap and still came up short (Energy Evolution, Aug 4). The read-through: chips can only ship as fast as customers can secure power, so on-site gas turbines and the equipment behind them are riding the same wave.

  • CPUs are a sneaky beneficiary of "agentic" AI. Several hosts made the same point: as AI shifts to autonomous "agents" that run many steps and tool-calls, the ordinary server CPU that orchestrates the GPUs becomes a bottleneck of its own. AMD said its server-chip revenue is set to grow more than 80% in the second half, and prices are up 35–40% (Elon Musk Podcast; TITV). The GPU gets the headlines; the CPU quietly gets pricing power.

  • "The whole pie is growing" is the emerging consensus. From Rasgon ("is the opportunity still bigger or is it not?") to Lisa Su's $1.3–1.4T market call to Baker's "under-earning" thesis, the bull case has quietly shifted from "who wins" to "the market is big enough for several winners." That's a subtle but important change in how the Street is framing Nvidia-vs-AMD-vs-custom-silicon.

What changed since Issue 012

  • The frame flipped from "one company on trial" to "the whole field being graded." Issue 012 was about Alphabet's cash-negative shock and a market that "punished a blowout." This week the other giants reported, and the market discriminated: Microsoft and Amazon rewarded for proven returns, Meta punished for unproven ones. And the "AI trade is over" fear from a week ago reversed violently, a ~$3.5 trillion, four-day rebound.

  • AMD's operator silence finally broke, decisively. For eight issues its executives were absent even amid big deals; this week AMD's earnings, Lisa Su's own words, the Helios "all five labs" milestone, and a $1.3–1.4T market call put it at the center of the story (even as the stock fell).

  • New heavyweights entered the bubble ring. Jim Chanos (bear) and Gavin Baker (bull) replaced last issue's Cuban / Chamath-Friedberg-Sacks voices, and Baker brought the most concrete new data point of the run: GPU rental prices rising 50–100%.

  • Memory gained operator confirmation. Last issue framed memory as "bear market on screen, shortage in the ground." This week Tim Cook ("100-year flood") and Andy Jassy (capex up on memory costs) confirmed the shortage is structural and multi-year, and the stocks began to recover.

  • Anthropic joined the custom-silicon club, and a fresh wave of deals (Volta/Norway, OpenAI's Georgia and Ohio sites) extended the "circular financing" web the bears keep circling.

  • Networking partly un-darkened via Arista's blowout, but pure optics stayed silent for another week.

  • Not re-run this week (negative space): Super Micro was quiet ahead of its August 11 report; TSMC and advanced-packaging (CoWoS) coverage was thin; Marvell and Broadcom drew only price chatter, not operator commentary; Meta's MTIA chip went unmentioned; and last issue's Cerebras/Feldman and Jensen-in-Korea threads did not recur.