Newsletter · · Ashutosh Agarwal

Nvidia Guides Higher and Cuts Margins on Memory Costs - AI Accelerators 019 - Week of August 27, 2026

GPUs, custom silicon and optics newsletter for the week of August 27, 2026, issue 019. Nvidia posted record revenue and raised its full-year outlook to roughly plus 70 percent, then cut its own gross-margin guide on extreme memory pricing and answered the circular-financing bears by name.

AI Accelerators

Week of August 27, 2026: Nvidia Guides Higher and Cuts Margins on Memory Costs

GPUs, Custom Silicon & Optics, a twice-weekly read on the hardware behind the AI build-out

Issue 019 · Thursday, August 27, 2026

Two days ago the framing here was that nobody was nervous about Nvidia's numbers; everybody was nervous about how the build-out gets paid for. Wednesday night, both halves of that got answered at once, and not in the way the bears spent all week rehearsing.

Nvidia reported after the close on August 26. Revenue was a record and the forward guide got bigger, not smaller. But for the first time in this cycle, management walked in and cut its own margin outlook, because the memory shortage the podcasts have been shouting about for a month is now showing up inside Nvidia's own cost line. And then, in the calmest possible voice, the CFO looked straight at the "circular financing" charge that has dominated this newsletter for three issues and said: we know what you call it, and we see it differently.

That timing is what makes this issue useful. The trailing week of podcasts is the exact bear case the market built into the print, and it can now be laid directly against what Nvidia actually said. When the argument and the rebuttal land in the same 24 hours, you get to grade both.

A quick glossary as we go, because the plain-English rule matters here: a hyperscaler is a giant cloud (Amazon, Microsoft, Google and the like); a neocloud is a newer company that does nothing but rent out AI computing (CoreWeave, Nebius); gross margin is the profit left after the cost of building the product; HBM is the ultra-fast "high-bandwidth memory" stacked next to an AI chip; an SPV is a special-purpose vehicle, a shell company set up to hold debt off the main company's books; take-or-pay is a contract where you pay for capacity whether or not you use it.

TL;DR

  • The number roared. Q2 revenue $96B, more than double a year ago. Data center $89B. Next quarter guided to $108B, and the full-year FY2028 outlook was raised to roughly +70%, which analyst Stacy Rasgon pegged as about $200B more than the prior plan.
  • The margin got cut, and memory is the reason. Nvidia reset gross margin down to ~74% next quarter, bottoming at 71–72%, on "extreme pricing conditions in memory." That is the podcasts' memory-shortage thesis (Howard Chan, Jim Handy) confirmed inside Nvidia's own P&L.
  • Nvidia answered the "circular financing" bears directly. CFO Colette Kress: "we know some will call this circular financing. We see it differently." Disclosed ~$50B into AI labs, $500B+ financing platforms with six Wall Street firms, and credit support for nearly 2 GW of compute for "another frontier AI lab."
  • The pre-print podcast week was almost entirely the bear case: Eurodollar University's Jeff Snider calling Nvidia "the gray swan" at the center of a house of cards; SharePickers' Justin Waite warning of a crash "three to five times bigger" than dot-com; the Broadcom debt deal above $60B spreading the same structure to a second chipmaker.
  • China is zeroed out. Less than 1% of data-center revenue went to China as H200 chips, and there is no China revenue at all in the forward guide.

1. The print: demand accelerated, and the guide got bigger

Start with what actually printed, straight from Wednesday's earnings call (Nvidia Q2 FY2027 earnings call, Aug 26, 2026). These are the company's own figures, and the podcasts had not seen them when they recorded their pre-print worry.

  • Total revenue $96 billion, "more than doubled year-over-year," with growth accelerating for a fourth straight quarter. CFO Colette Kress's word, "accelerated," matters, because the whole bear case rests on growth slowing.
  • Data center $89 billion, up 18% from the prior quarter. Kress split it two ways: hyperscale $49 billion (+13%, driven by Blackwell) and ACIE $40 billion, the neocloud, enterprise and sovereign bucket, up 25% sequentially and 138% year-over-year. That second bucket is the part Jensen Huang kept insisting the market cannot see: "There is an entire part of the market... sovereign AI, there are regional AIs, there are neoclouds, there are AI startups... which represents about half of our business, and that is growing 100% a year."
  • Networking set another record (+18%), with Spectrum-X Ethernet up 2.6x year-over-year, worth flagging because networking and optics have been dead air in the podcasts for weeks.
  • The guide is the headline. Next quarter (Q3) is guided to $108 billion, plus or minus 2%. And Nvidia did something it has never done, guiding a full year out: FY2028 revenue "to grow approximately 70% year-over-year," and Kress was explicit that this is supply-constrained: "customers' forecasts point to our growth doubling next year," meaning the unconstrained number is closer to +100%.

On the call, Bernstein's Stacy Rasgon, the analyst quoted here two days ago wanting "very, very good," did the arithmetic out loud: "The 70% growth in fiscal 2028... that is something like... a $200 billion uptick versus the prior outlook... The prior outlook was $1 trillion over the three years, so this is probably $200 billion more." Jensen's reply to what the unconstrained number would be: "The unconstrained would be a lot higher... We are just going to have to go work hard to get more capacity."

Two more things that move numbers:

  • China is gone from the model. Nvidia shipped "less than 1% of our total data center revenue in Hopper 200 products to customers based in China," those shipments are dilutive to margin, and, the key line, "there is no China data center compute revenue in our forward outlook." So the $108B guide and the +70% are built on zero China.
  • Capital returns hit a record. $26 billion back to shareholders in the quarter ($20B buybacks, $6B dividend), 60% of free cash flow year-to-date, with a stated intent to "increase and return excess free cash flow." That is not the cash profile of a company that thinks it is about to hit an air pocket.

Why it matters: the demand half of the pre-print debate resolved decisively in the bulls' favor. The bear case, to survive, now has to live entirely in the two places Nvidia itself flagged as pressure points, margins and financing. Which is exactly where the podcasts were pointing.

2. The surprise: Nvidia cut its own margins, and memory is why

This is the genuinely new, non-consensus development, and it is a beautiful piece of cross-confirmation.

For weeks the memory analysts on the podcasts have said the same thing: HBM is sold out, prices are ripping, and that has to land on someone's income statement. On August 23, Howard Chan of Kurv Investment Management laid it out on Behind the Ticker. His framing, that memory is not the picks and shovels but "the mines":

  • "90% of the memory chips in the world is actually made by three companies," Micron, Samsung and SK Hynix, and "they have sold out most of their inventories up to 2028."
  • "SK Hynix just reported earnings. Their revenue went up 250% and their profit went up 500%. So they are price setters instead of price takers."
  • New supply cannot rescue it quickly: a fab costs "30 to 50 billion" and years to build, and the lithography machines come from "only one company in the world... ASML... already have years of wait list." Chan calls it a "two to four-year thesis."
  • He even pre-empted the Chinese-competition worry: the new Chinese entrant is pricing "at the same level as the Korean company... not a cheaper alternative that people thought it was going to be."

Now put that against what Kress said Wednesday night. Nvidia held gross margin at 75% for the quarter just reported, but reset the forward outlook down:

"we are experiencing extreme pricing conditions in memory. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year. As a result, we are resetting expectations today."

The new shape: ~74% next quarter, bottoming at 71–72% in Q4, then settling at 72–73% for FY2028 once Nvidia's own price increases kick in. Recall the pre-print bar in Issue 018 was for Nvidia to hold the mid-70s. It did not, and it told everyone why, up front, rather than letting it dribble out. Kress framed the cut as a good problem: "Memory scarcity today is being driven in large part by the AI build-out itself... Tighter memory supply is a symptom of the same demand surge that is driving our own growth."

Read the two together and the takeaway is clean: the podcast memory bulls were right, and the proof just showed up in the P&L of their biggest customer. The pain is real (Nvidia is eating a few points of margin), but it is a demand symptom, not a demand warning. And it is a straight bullish read-through to the three memory names Chan named: the companies with the pricing power are the ones taking margin from Nvidia, not giving it.

(Operator versus pundit note: Chan runs a memory-focused fund, so treat his framing as an informed pundit talking his book. The confirming data point, the margin reset, is from Nvidia's own management on the record.)

3. The main event: the circular-financing bears got their week, and then Nvidia answered

Here is where the timing is almost too neat. The podcasts in the trailing week were, overwhelmingly, one long prosecution of Nvidia's financing. Then the defendant took the stand.

The prosecution. On August 21, Jeff Snider walked through the mechanics on Eurodollar University, and it is worth hearing him out, because it is the most technically careful bear case of the week:

  • The structure routes Wall Street money through an SPV that lends to Nvidia's customers, secured by the chips they buy. But the collateral is "relatively illiquid... specialized... and given the technology cycles, it's going to depreciate relatively quickly."
  • So Nvidia sweetens it with a residual value guarantee, a promise to make the SPV whole if the chips are worth less than the loan. And here is his sharpest point: that promise "doesn't show up on a balance sheet anywhere... just like credit default swaps a generation ago, there is in NVIDIA's perspective... almost no chance that these guarantees... are going to be triggered."
  • His historical rhyme is Nortel and Lucent, telecom-gear makers who vendor-financed their own customers into the dot-com boom and went bankrupt when the timing slipped: "It's not the technology that's bad or the technology that's wrong. More often than not, it's just the timing."
  • The verdict line: "NVIDIA sort of becomes the gray swan because it's at the middle of so many of these Jenga pieces... a house of cards where you start tugging one of the bottom row... and things look a lot different."

He was not alone. On The SharePickers Podcast (Aug 22), host Justin Waite ran the same play through CoreWeave's numbers: market cap ~$50B, enterprise value ~$100B once you add debt, and trailing free cash flow of negative $13.66 billion. His mechanism: Nvidia puts in ~$2.1B of equity plus preferred chip allocation, CoreWeave borrows against those chips to buy more chips, and the chips depreciate over six years while "NVIDIA are developing new chips pretty much every 12 to 18 months." His conclusion was the loudest of the week: "this is a house of cards that will come crashing down. And it will be worse than the dot-com bubble... you can expect a crash three to five times bigger."

And the structure kept spreading. The Broadcom debt deal flagged Monday got fresh reporting from Bloomberg News's Ryan Gould on Bloomberg Businessweek (Aug 20): Broadcom in talks to line up "as much as $100 billion in debt," roughly $30B junior from Apollo and Blackstone, with Broadcom itself "backstopping... somewhere in the region of $60 to $70 billion" of senior secured debt, extending the June Apollo and Blackstone partnership toward 20 gigawatts of compute, with Anthropic the lead tenant. Gould's own aside said the quiet part: "do we start to talk about the circular financing thing? Yes." (Figures are fluid and source-based, so treat them as reporting, not confirmed terms.) Separately, Bloomberg Tech reported Anthropic is prepping a blockbuster IPO, a detail that becomes relevant in a second.

The defense. Now here is Nvidia, Wednesday night, addressing the exact charge, not dodging it, naming it. Kress laid out the scale first: nearly $50 billion already invested in the frontier labs ("a small fraction of our expected free cash flow"); the $500 billion+ financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR; a SoftBank-hosted 4.25 GW campus in Portsmouth for OpenAI; OpenAI committed to ~12 GW through 2030; and "for another frontier AI lab, we will provide selective credit enhancement for nearly 2 GW of compute." Then, unprompted:

"We recognize the scale of this support, and we know some will call this circular financing. We see it differently. We are going through a major computing platform shift... these are once-in-a-generation companies... We expect them to become the largest technology companies in history."

The rebuttal has three pillars worth separating out:

  1. It is a slice, not the whole. The labs where Nvidia leans on its own balance sheet will be "roughly a quarter of our business next year," and the rest "will be consumed by investment-grade customers or those that are backed by one."
  2. The collateral is redeployable. Directly answering Snider's depreciation point: "The NVIDIA Compute platform is fungible and durable and can be redeployed to support other customers." Jensen went further, claiming ROIC on a $50B data center is now "less than a year."
  3. A new twist, Nvidia gets paid twice. For neoclouds, Nvidia now offers a take-or-pay minimum-revenue guarantee so lenders will underwrite the project, "and in exchange, we share in a portion of the neocloud's revenue earned above that floor... We're not making loans." Kress: "we get paid twice, once on the hardware sale and again through the share of rental revenue."

And when BofA's Vivek Arya added up all the commitments to "about $500 billion" and asked if that was the ceiling, Jensen did not flinch: "the only regret that I have is that I didn't invest more and sooner. Two of the companies will likely go public soon." Which is where the Anthropic IPO reporting suddenly fits: the labs Nvidia is backstopping are heading for public markets, which is exactly the "investment-grade or backed by one" story management is selling.

Why it matters: this is the cleanest bull-versus-bear collision the sector has produced. The bears (Snider, Waite) say the residual-value guarantees are hidden CDS-style risk and the collateral is a depreciating asset in a customer base that cannot pay if demand slips. Nvidia says the guarantees are a small, fungible slice against once-in-a-generation demand, and it now monetizes the rental stream on top. The swing factor is unchanged and unfalsifiable until it breaks: does demand hold long enough for the debt to get repaid? Snider's Nortel line, "it's just the timing," and Jensen's "less than a year" ROIC are the two ends of that exact bet. What did change Wednesday is that Nvidia stopped letting the question "linger as an open question" (Kress's phrase about margins, but it applies to the whole call) and argued the case out loud.

(The standing operator rebuttal to the depreciation bears remains the CoreWeave co-founder on The Rundown from last week, 2020-vintage A100s still on lease at full price and self-amortizing contracts, which is the operator version of Jensen's "fungible and durable" claim. Nvidia's management this week is the only fresh operator voice; the financing bears are all pundits.)

4. Custom silicon: the only place it showed up was Nvidia's own call

There were no fresh podcasts this week on Google's TPU, Amazon's Trainium, Microsoft's Maia or Meta's MTIA. The only custom-silicon debate happened on Nvidia's call, when Vivek Arya pointed out that the labs Nvidia is financing, OpenAI and Anthropic, are building their own chips, citing OpenAI's "Jalapeño" and "their claims about being better than Blackwell."

Jensen's rebuttal is the one to file for when Marvell and Broadcom report:

"Whereas many of these XPUs are inference-specific chips for one cloud or one service, NVIDIA is a platform, an entire AI factory platform that spans the entire AI life cycle that you can use in any cloud."

He also quietly widened the platform: Nvidia's revenue opportunity per gigawatt of data center has gone from ~$18B (Hopper) to $25B (Blackwell) to $40B (Vera Rubin), now spanning the Vera CPU (in full production, shipping to Oracle, SpaceX's AI unit and AWS), three kinds of networking, and the Groq 3 LPX rack-scale "LPU" for ultra-low-latency inference (in production, shipping in volume this quarter, Nebius first). The message to the ASIC bulls: Nvidia is trying to sell you the whole factory, not a chip. Whether the custom-silicon crowd has a counter is a Marvell and Broadcom question.

5. Optics: a flicker of life, mostly a rebound story

The one optics episode of the week was The MoneyFlows Show (Aug 20), and it was more of a bounce-back trade than fresh fundamentals: the photonics names had sold off during the forced liquidation of Leopold Aschenbrenner's Situational Awareness fund, and the hosts framed the dip as an entry point. Still, a few concrete data points worth keeping:

  • Coherent (COHR), "lasers that make the light," silicon photonics and transceivers now pushing 3.2 terabits per second, with a $2 billion strategic investment from Nvidia; backlog extending beyond 2028.
  • Lumentum (LITE), lasers, modulators and silicon photonics; the hosts called it "one of the best earnings reports of the season."
  • Fabrinet (FN), the precision manufacturer, targeting a $12.5–14B revenue run-rate over the next several years.
  • Tower Semiconductor (TSEM), the silicon-photonics foundry, guiding to a $1B annualized silicon-photonics run-rate by Q4 2026, and already supplying 1.6-terabit modules for Nvidia's networking.

This is a pundit-only, momentum-flavored take with no operator on the mic, so treat the specific price-target enthusiasm accordingly. But it lines up with the one hard fact from Nvidia's call: networking grew 18% and Spectrum-X is up 2.6x. The optical supply chain is real; the podcasts just are not covering it with much depth right now.

6. Names in play and read-throughs

  • Marvell (MRVL) reports tonight (Aug 27, after close). This corrects Monday's forward note, which had it on the 25th. It is the first custom-silicon print after Nvidia planted its "we sell the whole factory, not a chip" flag, and it comes off the $12.2B Google TPU warrant. Watch data-center growth and any new custom-chip customers; that is the direct test of Jensen's platform-versus-ASIC argument.
  • Memory (Micron, SK Hynix, Samsung). Nvidia just confirmed memory pricing is "headed even higher into next year" and is eating a few points of its own margin to prove it. That is a straight positive read-through to the three suppliers with the pricing power. Micron is the US-listed way to play it.
  • AWS and hyperscaler capex. Nvidia disclosed AWS is deploying an additional 2 million GPUs through early 2029 plus Vera CPUs, and put top-five hyperscaler capex at ~$800B in 2026 and $1.3 trillion in 2027. The capex spigot is still opening, not closing.
  • Neoclouds (CoreWeave, Nebius). Two opposing reads collided this week: SharePickers' negative $13.66B free-cash-flow bear math versus Nvidia's new "we get paid twice" revenue share and the disclosure that neoclouds exit the year at 8 GW, up from ~3 GW. Nebius also gets first dibs on the Groq 3 LPX. This is the highest-beta expression of the whole debate.
  • Broadcom (AVGO). The debt deal above $60B (up to ~$100B) is the financing structure jumping to a second chipmaker. If Nvidia's circular-financing defense holds up in the market, it de-risks Broadcom's version too; if it cracks, both correlate.

7. What changed since Issue 018

  • The frame flipped from "on the clock" to "off the clock." Monday was the last preview; the bar was Rasgon's "very, very good," ~$92–95B for the quarter, hold the mid-70s margin, and give financing detail. Grading it: revenue beat ($96B), guide raised ($108B and +70% FY28), margin bar missed (cut to 71–73%), financing detail delivered (in full, and on offense).
  • Memory went from the podcasts' loudest thesis to a line in Nvidia's guidance. Monday it was Howard Chan and Jim Handy saying HBM is sold out; Wednesday it was Nvidia cutting margin because of it.
  • The financing worry that spread to Broadcom on Monday got its rebuttal. Nvidia named "circular financing" and argued against it, the first time the accused has spoken since this became the franchise's spine.
  • China moved from "assume zero" to officially zero in the forward model.

Thin or silent this week: no fresh podcasts on Google TPU, AWS Trainium, Microsoft Maia or Meta MTIA, and nothing on TSMC or CoWoS packaging, so the custom-silicon and foundry conversation happened only inside Nvidia's call. AMD is quiet for a third straight issue after the flood around its own print earlier this month. Optics is one momentum episode deep against a real supply chain.

The tell across all of it: the podcast ecosystem spent the week almost entirely on risk, meaning debt, depreciation, circularity, and "five times worse than dot-com." Then the company at the center of it put up an accelerating number, raised the year, and answered the risk case point by point. The gap between that mood and that print is the single most interesting thing to watch from here.