Nvidia cleared every bar it set for itself on August 26 and the stock went down anyway. Revenue of $96.22 billion, up 105.8% year over year. Data center at $89.02 billion, up 117%. A third-quarter guide of $108 billion against a $104.2 billion consensus, with China data center revenue assumed at zero. Adjusted EPS of $2.22 against $2.09. The shares slipped, and they remain below the $236.54 record set in May.
The line that did the damage was further down. Colette Kress guided gross margin to 74% for the current quarter, said it would trough at 71% to 72% in the January quarter, and volunteered the reason rather than waiting to be asked about it: memory scarcity, driven in large part by the AI buildout itself.
That is the frame for the week of September 1. Nvidia is the most advantaged buyer of high-bandwidth memory on the planet. It has the volume, the roadmap visibility, the multi-year supply agreements, and a 75% gross margin to absorb a shock with. It is still handing three to four points of gross margin to its own suppliers over the next two quarters. Everyone reporting between Tuesday and Thursday sits downstream of that same bill of materials with far less cushion, and the useful question is not whether their AI revenue grew. It grew. The question is what it cost them, and whether they can push the cost forward.
Dell reports Tuesday with the thinnest structure in the group. The company runs somewhere around 18% to 19% gross margin at the consolidated level, which means a two-point move in component cost is a catastrophe rather than a talking point. Consensus sits near $44.9 billion in revenue and roughly $4.89 in earnings, against management’s own $44 billion to $45 billion range. The interesting number is elsewhere. Dell booked $24.4 billion of AI orders last quarter, recognized $16.1 billion of AI server revenue, and exited with a record $51.3 billion backlog. Guidance for the quarter just ended calls for approximately $15.5 billion of AI server revenue. Record backlog, and sequentially lower shipments. Evercore’s read is that supply constraints now span memory, CPUs, optical components and hard drives at once, and that the pipeline is several times the backlog. Both things can be true, and if they are, the backlog stops being a bullish signal and becomes a measurement of how much revenue Dell cannot physically deliver. Watch ISG operating margin, not the ISG revenue line.
Broadcom reports Wednesday into the hardest setup on the calendar. The company guided third-quarter revenue to $29.4 billion and AI semiconductor revenue to $16 billion, more than 200% growth and over half of the total. Consensus is roughly $29.36 billion and $3.24. Management has already flagged that gross margin steps down to about 74% on product mix, which is the custom accelerator business diluting the networking business as it scales, the same shape Marvell showed on August 27 when it guided incremental gross margin near 52% against 58.9% delivered. Broadcom sold off 12.6% after its last beat because Hock Tan declined to raise the fiscal 2027 AI target above $100 billion. Whether he raises it now is the swing factor. There are two other disclosures worth more than the headline. The first is the quarterly maximum exposure figure on the Apollo and Blackstone AI XPV transaction, which stood at $29 billion against an initial $35 billion structure in the last 10-Q; if that number stays roughly flat, the backstop is a launch subsidy, and if it scales with gigawatts, the market has been pricing the wrong denominator. The second is Google. Marvell’s August warrant gave Google the right to acquire up to $12.2 billion of stock against custom silicon revenue, and Broadcom has been Google’s primary TPU partner. Nobody is going to say anything useful about share loss on the call. The revenue mix inside the AI segment will say it for them.
Credo reports Tuesday and is the cleanest test of the conversion problem, because management has already published the contradiction. Guidance calls for $465 million to $475 million, growth of 108% to 113% against $223.1 million a year ago, on non-GAAP gross margin of 67% to 69%. Consensus is $470.7 million and $1.16. For the full fiscal year, management is guiding to more than 80% growth with over $600 million of optical revenue, and ZeroFlap optics, silicon photonics and optical DSPs each contributing more than $100 million. But it has also guided mid-single-digit sequential growth through the first half. Mid-single-digit sequential growth in two quarters, and 80% for the year, means an extraordinary second half, gated on optical components that are the tightest part of the supply chain. Credo fell 9% to 14% after the fiscal 2026 print on exactly this concern, and the DustPhotonics acquisition at $750 million was the answer: own the photonic integrated circuit rather than queue for it. If the vertical integration is working, it shows up in first-half gross margin holding the range, not in the annual guide being reaffirmed.
Below the headline names the same test repeats with worse leverage. NetApp reports Wednesday with options pricing a possible double-digit move, and its cost of goods is the NAND and hard drive market that has been repricing all year. Ciena reports Thursday into optical demand that exceeds supply and coherent DSP lead times that have been extending since spring. Ambarella reports Thursday guiding $105 million to $111 million on 59% to 60.5% gross margin, which is the lowest margin structure in the group and the weakest negotiating position at the foundry, with 42 million AI SoC units deployed and a partnership announced at over $800 million of potential lifetime revenue. Small companies buy wafers last.
The software half of the week is usually treated as a separate conversation. It should not be. Snowflake and MongoDB sell consumption, and consumption is compute, and compute is the thing whose input costs Nvidia just described as scarce. The pass-through question is identical to Dell’s, only the customer is an enterprise data team rather than a hyperscaler procurement desk. Snowflake reports Wednesday with product revenue growing near 34% and a five-year, $6 billion AWS commitment behind it, which is a rate lock as much as a partnership. MongoDB reports Tuesday. Both go into the print after the bar moved violently the week before, when Salesforce rose about 20% on Agentforce reaching $1.5 billion of ARR, Okta jumped on AI identity products driving roughly 30% of new bookings, and CrowdStrike, Workday, Atlassian and ServiceNow all rallied. The SaaS-is-dead trade was unwound in five sessions. GitLab, UiPath and DocuSign report into the aftermath with the least defensible unit of billing in the group, since each sells something an agent can plausibly consume rather than purchase: seats, robots, envelopes. Palo Alto Networks on Tuesday and Zscaler on Thursday carry the security read, and security is the one software category where AI expands the addressable problem faster than it automates the vendor away.
Which brings the week down to pricing power, which is the only durable form of moat in an input-cost shock. Broadcom has it, because networking attach and custom sockets carry switching costs measured in silicon respins. Credo has it if DustPhotonics closes the loop on photonics and does not if it only shortens the queue. Snowflake and MongoDB have it structurally, because consumption pricing reprices continuously, but only to the extent customers keep consuming at a higher unit price. Dell has scale procurement and almost nothing else, which is why it converts the most revenue and keeps the least of it. Ambarella has the least of anyone here.
The positioning does not help. Dell is up more than 250% year to date. Credo trades near $283 after a 218% twelve-month run. Broadcom sits around $354, below its twenty, fifty and two-hundred day averages, with an average street target above $518, which is a spread that describes disagreement rather than upside. Marvell raised guidance on August 27 and fell. Nvidia beat and guided above consensus and fell. Three of the last four AI infrastructure prints have been sold into. A beat is not a catalyst in this tape; a margin trough that gets named and dated might be.
Nvidia told the market where its bottom is and when it arrives. Nobody reporting next week has done that. The disclosure worth waiting for is not the AI revenue number any of these companies put on the first page. It is whether a single one of them will say out loud, as Kress did, which quarter the gross margin stops going down.
Leave a Reply