Elon Musk’s exit from AMD chips grabbed the headline. The three customers who stayed, and what AMD paid to keep them, are the actual story.
Advanced Micro Devices reported record revenue and a clean top-and-bottom-line beat on August 4, then dropped about 8% after hours the moment Elon Musk told SpaceX investors he would buy only Nvidia going forward. Here is the part the selloff skipped: SpaceX had close to zero gigawatts of AMD chips on order, while OpenAI, Meta, and Anthropic have committed to roughly 14. The market punished the departure of the one customer AMD’s business model was never built to serve, and mostly ignored the vendor financing that makes its real backlog look bigger than it is.
Record quarters are supposed to be good news. This one wiped out tens of billions in market value in a single after-hours session, and the trigger was a rival CEO praising a competitor on a different earnings call. To understand why a company can beat every number and still get hit, you have to stop reading the headline and start reading the customer list.
What Happened
AMD booked $11.54 billion in revenue for the quarter ended June 27, up 50% year over year and ahead of the roughly $11.3 billion analysts expected. Adjusted earnings came in at $1.66 a share against a $1.61 estimate. Data center revenue more than doubled to $6.7 billion and now makes up about 58% of the company, up from 42% a year ago. Gross margin expanded to 56%. Management guided third-quarter revenue to around $13 billion. On almost every line, AMD did what it needed to do.
The stock had run up 7% during the day to close at $518.58, already up more than 130% on the year. Then two things landed after the bell. AMD’s own capital spending had doubled from a year earlier, which spooked a market already nervous about AI overspending. And on SpaceX’s earnings call happening at almost the same time, Musk said his company would build exclusively on Nvidia because he considers its architecture the best available. As recently as May, Musk had said SpaceX and Tesla would likely keep buying from both suppliers. The reversal was enough to send AMD down close to 9% in extended trading, to about $473.
Note the scale of what actually moved the stock. The customer whose exit made the headline had ordered nothing. The customers keeping AMD’s data center engine running had ordered more compute than most countries can power.
The Backstory
AMD spent the last year turning itself from Nvidia’s distant runner-up into the default second source for anyone building frontier AI. In October 2025 it signed OpenAI to deploy up to 6 gigawatts of Instinct GPUs, and sweetened it with a warrant that lets OpenAI buy up to 160 million AMD shares, close to a tenth of the company. It lined up a 6-gigawatt partnership with Meta. In July it announced Anthropic would deploy up to 2 gigawatts of its MI450 chips inside AMD’s new server racks, alongside a commitment from AMD to invest up to $5 billion into Anthropic itself. Microsoft agreed to run AMD’s racks on Azure. That is the customer book the market was too distracted to price.
SpaceX was never on that list in any meaningful size. Musk’s companies build their own everything, from rockets to the xAI compute cluster, and a vertically integrated buyer chasing maximum performance has no reason to hedge across two chip vendors. Which is exactly why his exit says less about AMD than it seems to.
The Plan
AMD is no longer trying to sell you a faster chip. It is trying to sell you the whole rack. The centerpiece is Helios, a rack-scale system built to specifications drawn up with Meta and aimed squarely at Nvidia’s NVL72. Around it sit the MI450 GPUs, sixth-generation EPYC server processors, Pensando networking, and the ROCm software layer, all pitched on a single metric AMD keeps repeating: inference tokens per dollar. The company raised its forecast for the data center AI chip market to roughly $1.4 trillion by 2030 and expects its server CPU business to grow more than 80% in the second half of this year.
The strategy is coherent. The buyers running inference at scale want a credible alternative to Nvidia so they are not captive to one supplier’s pricing and roadmap. AMD’s plan is to be that alternative across the entire stack, not just on one chip. Being the hedge is the product.
The Business Model Angle
Here is the reframe the selloff missed. AMD’s role in the AI economy is to sell insurance against Nvidia lock-in. Its addressable market is not “everyone who needs GPUs.” It is the narrower set of buyers who are large enough to fear single-vendor dependence and independent enough to want a second option. OpenAI, Meta, Anthropic, and Microsoft are all in that set, and all four re-upped. Musk is not. He is building a closed vertical stack and optimizing for one thing, so he wants the single best architecture and does not value the hedge. His exit is the exception that defines AMD’s market, not a verdict on it. The market sold the stock on the one buyer who was never really a customer.
Now the uncomfortable half. AMD’s real backlog is not as clean as “record demand” implies, because AMD is helping pay for it. The OpenAI warrant hands over close to 10% of the company, vesting as OpenAI deploys chips and as AMD’s stock climbs toward a $600 target. The Anthropic deal comes with up to $5 billion of AMD’s own cash flowing into the customer. This is the same circular pattern reshaping the whole sector, where the profit pool keeps migrating around the AI stack and suppliers increasingly fund their own buyers. Demand you paid for is softer than demand that shows up on its own, and it dilutes you precisely when things go well.
The Risk
The biggest risk is not Musk. It is that AMD’s diversification is partly an illusion. OpenAI, Meta, and Anthropic look like three different customers, but they are one trade wearing three logos. All three are levered to the same AI capital-spending cycle and the same unanswered question of whether AI revenue will ever catch up to AI capex. In July, hardware stocks sold off hard on fears that AI companies had overextended and might not meet their financing commitments. If that fear returns, AMD’s three anchor customers wobble together, not separately.
The vendor financing compounds it. The OpenAI warrant dilutes AMD shareholders exactly as the stock works, and the $5 billion Anthropic check is capital AMD spends to secure revenue it then books as a win. Timing makes it sharper: the first big deployments are a late-2026 and 2027 story, yet the stock trades near 50 times forward earnings today. A market priced for perfection punishes narrow beats, which is part of what just happened. And multi-sourcing cuts both ways. Anthropic buys Nvidia too, and Nvidia has committed up to $10 billion to Anthropic in return. The same logic that keeps AMD’s customers from being locked into Nvidia keeps them from being locked into AMD.
To be fair to the sellers, there is a real counterargument. Maybe Musk’s comment is genuine product information. If the most compute-hungry, technically fluent buyer in the market looks at both and picks Nvidia on merit, that is a data point about AMD’s chips, not just about Musk’s vertical integration. And the quarter did carry soft spots, including operating income that came in slightly below expectations. Musk may have been the trigger rather than the cause, but a stock this richly valued did not need much of a reason.
Quick Questions
Did AMD have a bad quarter? No. It was a record, with revenue up 50% and data center sales more than doubling. The stock fell on sentiment and valuation, not on the results.
How much did losing SpaceX actually cost AMD? Very little in the near term. SpaceX was a small AI buyer for AMD, and analysts flagged the announcement as having minimal material impact. The reaction was about Musk’s influence on sentiment, not lost revenue.
Why is AMD investing in its own customers? To lock in multi-year, multi-gigawatt demand in a market where every hyperscaler is courted by Nvidia. It works as a demand signal, but it blurs the line between a sale and a subsidy.
Is AMD a real threat to Nvidia now? As a credible second source, yes. As the preferred architecture, not yet. Its pitch is choice and tokens per dollar, not raw performance leadership.
The Business Model Analyst Take
The headline said a customer left. The business model says the opposite: AMD just proved it can sign the buyers its whole strategy depends on, and the one who walked was never in the market it serves. That is the bull case, and it is stronger than an 8% drop suggests.
The bear case is quieter and more serious. AMD’s record book is concentrated in three frontier labs riding a single capital-spending wave, and AMD is financing some of that demand with its own equity and cash. Sturdy businesses do not usually have to pay their customers to show up. The number to watch is not whether Musk comes back. It is whether OpenAI, Meta, and Anthropic keep spending when the AI capex debate finally forces someone to justify the bill. If they do, AMD grows into its multiple. If they blink, they blink together, and the diversification that looks like a moat today turns out to be one bet with three names on it.
