AMD reported the best quarterly earnings in its 56-year history Tuesday evening — record revenue, doubled data center sales, and guidance that topped analyst estimates — and watched its stock drop more than 8% anyway, because Elon Musk used SpaceX's first earnings call as a public forum to declare his companies would be "exclusive to Nvidia" for all future AI infrastructure, on Earth and in orbit. The AMD Q2 2026 earnings press release confirmed the figures that made the drop all the more striking.

The declaration, delivered Tuesday evening during SpaceX's inaugural quarterly report as a public company, landed as a reminder of a structural reality that AMD's financial momentum has not yet dislodged: in the AI GPU market, a single high-profile endorsement from a marquee customer moves markets more than a record income statement. AMD's stock fell 8.94% in after-hours trading to $472.20. Nvidia gained roughly 2%.

The timing was particularly stark. Hours before the SpaceX call opened, AMD had reported second-quarter revenue of $11.5 billion — a company record, up 50% year-over-year — with data center revenue more than doubling to $6.72 billion, up 107% from the same period in 2025. Non-GAAP earnings came in at $1.66 per share against a $1.62 analyst estimate. For the third quarter, AMD guided for approximately $13 billion in revenue — above the $12.5 billion consensus. By any conventional measure, it was a blowout.

None of it survived the evening. CNBC's Jim Cramer wrote on X: "If Elon hadn't gone all in [on] Nvidia with what could be the biggest set of orders in the history of Nvidia, then AMD's stock would have been up last night."

Musk's Endorsement and What AMD Had Actually Posted

"Going forward, we have decided to build exclusively on Nvidia, because we think the Vera Rubin architecture is the best architecture," Musk said on the call. "We think it is the best AI computer, and we greatly value our close cooperation and partnership on many levels with Nvidia. We are exclusive to Nvidia." That SpaceX exclusive Nvidia pledge, reported by Tom's Hardware, wiped out AMD's after-hours gains before most investors had finished reading the earnings release.

AMD CEO Lisa Su pushed back from the company's own earnings call. "We are still in the early innings of a multiyear AI adoption cycle and the opportunity ahead is enormous," Su said in AMD's official earnings statement. "We are exceptionally well positioned to capitalize on this opportunity and deliver significant growth in the coming years." CFO Jean Hu added that data center sales are expected to accelerate further in the second half of 2026.

Starmind Will Carry Full NVL72 Racks Into Orbit

The most technically significant element of Musk's declaration was not the exclusive branding — xAI has run its entire Grok AI model training history on Nvidia hardware and has never deployed AMD Instinct chips in production — but the orbital deployment plan.

Musk confirmed that SpaceX and Nvidia are jointly developing the compute payload for Starmind AI1, a satellite designed to run large-scale AI workloads from low Earth orbit. Starmind satellites will each carry Nvidia's Rubin GPUs and Vera CPUs — the same silicon that powers Nvidia's flagship Vera Rubin NVL72 rack on the ground — in an "optimized" orbital configuration. "With respect to the Starmind AI satellite, which will be essentially an optimized Vera Rubin NVL72 computer, this is not some sort of far future distant thing; we expect to start launching this next year," Musk told investors on the call.

SpaceX plans to launch two prototype AI1 satellites in early 2027, with a commercial constellation to follow. For context on the engineering challenges that satellite orbital data centers face, see TechTimes' AI1 satellite analysis.

Why Cable-Free Design Makes the NVL72 Suited for Space

What makes the Vera Rubin NVL72 a credible orbital compute candidate — and what Musk's preference for it over "standard rack-style design" actually refers to — is an architectural decision Nvidia made for terrestrial deployment that has unexpected advantages in space.

Previous NVL72 generations (the GB200 and GB300 systems) required thousands of internal copper cables per rack. Each cable connection was a manual installation point and a potential failure. Assembly time for a single compute tray ran approximately 100 minutes.

The Vera Rubin NVL72 replaces all of that with a midplane-based design. Compute trays insert via blind-mate connectors into a single large printed circuit board midplane that spans the rack — no internal cable routing, no fans inside compute trays. Assembly time drops to five minutes, an 18-fold improvement. Jensen Huang highlighted this design change at his CES 2026 keynote as a core commercial advantage.

In orbit, those same properties matter for different reasons. Space deployment eliminates the option of manual cable routing after launch. Reducing mechanical failure vectors — fan bearings, connector seating errors, cable harness vibration fatigue — is not a convenience feature in a satellite; it is a mission-duration requirement. The cable-free NVL72 architecture, designed to simplify data-center operations, happens to align with what orbital hardware needs.

The scale of the Starmind ambition adds a further dimension. Nvidia's own SPACE-1 Vera Rubin Module, developed for orbital partners including Axiom Space and Starcloud, is designed to carry just a handful of Rubin accelerators per unit. Starmind satellites are planned to carry full NVL72-equivalent systems: 36 Vera CPUs and 72 Rubin GPUs per satellite, connected via NVLink 6's 260 terabytes-per-second all-to-all fabric. That is six times Nvidia's SPACE-1 GPU count.

The engineering challenges are real. The full NVL72 system generates substantial thermal load — well above the 150-kilowatt design documented for the first AI1 prototype. In space, waste heat cannot be shed into air or water; it can only radiate as infrared light from dedicated panels. Whether SpaceX's liquid radiator design scales to NVL72-class thermal output is an open question. Musk acknowledged as much when he said launching and connecting the hardware "could prove more difficult than constructing facilities on Earth."

Why AMD's Record Numbers Couldn't Compete With the Narrative

The mechanics of why AMD fell despite a record quarter reveal something specific about how the AI GPU market works that financial statements alone don't capture.

AMD holds roughly 5 to 7 percent of the AI accelerator market by revenue. Nvidia holds approximately 80 to 88 percent, depending on measurement methodology and whether custom ASICs are included in the denominator. A prior TechTimes AMD market share analysis documented this market structure in detail. Every high-profile customer win — or public rejection — carries outsized symbolic weight when the challenger's share is that small.

AMD has been accumulating substantial customer commitments. OpenAI has a six-gigawatt agreement for Helios rack deployments. Meta has a six-gigawatt commitment for AMD Instinct chips. Anthropic announced a two-gigawatt Helios deal alongside AMD's $5 billion equity investment. Microsoft has committed to deploying Helios on Azure. These are enormous gigawatt-scale orders, according to Fierce Network's analysis. On paper, they suggest AMD is winning hyperscale business at a pace that should shift the market-share narrative.

But none of those commitments produced a CEO from a marquee AI company standing up on an earnings call and calling AMD's platform the best. Musk did that for Nvidia. The CUDA software ecosystem — 18 years of optimized libraries, millions of developer-years of accumulated code, and deployment tooling deeply embedded in cloud platforms — has created a psychological and practical switching cost that hardware specs alone don't overcome. AMD's own ROCm platform has been narrowing the gap, but as TechTimes' AMD ROCm deep dive found, the ecosystem distance is structural.

Helios is a technically competitive system. AMD claims 30% more tokens per dollar than Nvidia's Rubin NVL72 system, and the open-standard UALink/Ethernet interconnect offers hyperscalers an alternative to NVLink lock-in. Those advantages matter to procurement teams running detailed cost models. They do not generate the kind of public declaration that moves a stock 8% in a single evening.

SpaceX's Own Stock Gained on Earnings Day

SpaceX's Q2 report was the strongest financial case the company has made since its June debut. Revenue came in at $7.81 billion against a $6.93 billion analyst consensus, and the AI Infrastructure segment produced $2.56 billion against a $2.18 billion estimate — a 312% sequential jump from $818 million in Q1, confirming that Anthropic's $1.25 billion-per-month compute lease is being recognized at nearly full rate. As TechTimes' same-day SpaceX Q2 analysis noted, management issued formal full-year revenue guidance for the first time in the company's 24-year history.

SpaceX stock (Nasdaq: SPCX) gained 9.4% during regular trading on August 4, closing at $125.33 — though the stock remained below its $135 IPO price. The most consequential event for SPCX shareholders will come Thursday, August 6, when approximately 911.5 million previously restricted insider shares are scheduled to become eligible for sale in the largest first-tranche lock-up expiry in stock market history.

---

Frequently Asked Questions

Why did AMD's stock fall 8% despite posting a record quarter?

Musk declared on SpaceX's earnings call that SpaceX and xAI would be "exclusive to Nvidia" going forward, citing the Vera Rubin architecture as the best available AI computing platform. AMD holds only about 5 to 7 percent of the AI accelerator market by revenue, which means high-profile customer signals carry disproportionate weight on its stock. The market read the Musk statement as a symbolic ceiling on AMD's momentum — even as AMD's actual financial trajectory (50% revenue growth, data center revenue more than doubling) was pointing sharply upward. The same dynamic works in reverse for Nvidia's roughly 2% gain on the news, as the after-hours market data confirmed.

What is Starmind, and how does it differ from Nvidia's own orbital compute plans?

Starmind is SpaceX's planned fleet of AI satellites, each designed to carry a full-scale equivalent of Nvidia's Vera Rubin NVL72 rack — 36 Vera CPUs and 72 Rubin GPUs — in low Earth orbit. SpaceX plans to begin launching prototype Starmind AI1 satellites in 2027. Nvidia's own orbital module, SPACE-1 Vera Rubin, is designed to carry several to a dozen Rubin AI accelerators per unit — a fraction of the Starmind scale. The key engineering question is thermal: deploying NVL72-level power loads in vacuum, where heat can only escape via radiative panels, is a materially harder problem than Nvidia has publicly addressed in its SPACE-1 specification, as TechTimes' prior AI1 analysis documented.

Why does the NVL72's cable-free design matter for space deployment?

Nvidia's Vera Rubin NVL72 replaced thousands of internal copper cables with a midplane printed circuit board, cutting compute tray assembly time from approximately 100 minutes to five minutes. In space, eliminating manual cable connections eliminates failure modes that are unserviceable after launch: cable seating errors, vibration fatigue, and fan bearing failures. The cable-free architecture was designed to simplify terrestrial data-center operations — its orbital relevance is a secondary benefit of the same engineering decision. Musk's stated preference for the NVL72 design over "standard rack-style design" directly references this midplane architecture.

Does AMD's pipeline of gigawatt-scale customer commitments change its competitive position against Nvidia?

AMD has secured substantial commitments: a six-gigawatt deal with OpenAI, a six-gigawatt deal with Meta, a two-gigawatt deal with Anthropic alongside a $5 billion AMD equity investment, and deployments with Microsoft on Azure and Oracle. These represent real infrastructure demand and signal that the largest AI buyers want an alternative to single-vendor dependence on Nvidia. But commitments and deployments are different things, and the CUDA ecosystem's 18-year depth means the software switching cost for training workloads remains a structural barrier that hardware specs and financial commitments haven't yet fully addressed. AMD's own ROCm platform has been narrowing the gap, but the symbolic weight of the Musk declaration shows how much brand narrative still shapes market perception independently of the underlying business fundamentals. See TechTimes' ROCm deep dive for a detailed assessment of where the software gap stands. This is not investment advice.