Equity for Silicon: How AMD Turned OpenAI and Meta Into Stakeholders

For most of the generative-AI buildout, the accelerator story was a one-vendor narrative. Hyperscalers and frontier labs competed for Nvidia allocation; everyone else competed to look like a credible alternative. AMD’s third-quarter 2026 results, released on 3 November, mark a different moment: named customers are already running production workloads on AMD’s new rack-scale systems, Data Center revenue nearly doubled year over year, and the cost of buying that credibility — warrants that turn chip buyers into potential shareholders — is now visible on the income statement.

This piece maps what AMD reported, what the OpenAI and Meta warrants actually say in the filings, and where contractual commitments end and projections begin. For the broader spend context that those customers sit inside, see our earlier map of hyperscaler AI capex.

The quarter that made the second source real

AMD reported third-quarter 2026 revenue of $13.4 billion, up 45% year over year. Gross margin was 54% on a GAAP basis (56% non-GAAP). Operating income was $2.8 billion GAAP and $3.9 billion non-GAAP; net income was $2.4 billion GAAP and $3.5 billion non-GAAP; diluted EPS was $1.46 GAAP and $2.09 non-GAAP.

The segment that matters for AI Power Players is Data Center. It delivered a record $8.5 billion, up 97% year over year and up 27% sequentially, and accounted for about 64% of company revenue in the quarter, according to AMD’s prepared remarks. Management attributed the surge to strong demand for EPYC server CPUs and the launch of Instinct MI450 Series GPUs.

Client and Gaming revenue was $3.7 billion, down 8% year over year. Embedded was $1.1 billion, up 29%. The mix shift is the story: AMD is becoming, quarter by quarter, a data-center AI company that still sells PCs and consoles on the side.

For the fourth quarter of 2026, AMD guided to revenue of approximately $17 billion, plus or minus 3%, with non-GAAP gross margin of about 55%. At the midpoint, that would be roughly 66% year-over-year growth and about 27% sequential growth. Chair and CEO Lisa Su said the company expects growth to accelerate further in the fourth quarter “as we scale Helios and ‘Venice’ production.”

Helios: from roadmap slide to named fleets

The product that turns “second source” from aspiration into procurement is Helios, AMD’s rack-scale AI system built around Instinct MI450 Series GPUs and 6th Gen EPYC CPUs. In the earnings release, AMD said Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure “are running models and workloads on the first AMD Helios rackscale solutions.”

That sentence is carefully present-tense and carefully limited. It does not claim that these customers have finished migrating training fleets, that Helios is their primary platform, or that committed dollar volumes have fully shipped. It says first systems are live and running real work. Prepared remarks add that AMD began shipping MI455X in volume in the quarter, that lead customers brought their first Helios racks online quickly, and that Helios deployments are expected to increase substantially in the fourth quarter as HPE, Dell, Lenovo, Supermicro and other OEM/ODM partners begin shipping systems.

HPE separately announced that Vultr will deploy the first AMD Helios by HPE systems. Oracle said its OCI Compute E7 instances will be powered by 6th Gen EPYC CPUs. AMD also introduced Helion, a reference architecture for rack-scale agentic systems, and released ROCm 10. Those are platform plays — necessary if AMD wants software lock-in to follow silicon share — but they are not yet the same as disclosed multi-year purchase minimums in a customer’s SEC filing.

The warrants: equity as a demand instrument

The financial twist is older than the Helios ramp, but Q3 is when it hit earnings hard. AMD’s Form 10-Q for the period ended 26 September 2026 states that the company issued warrants to two customers in October 2025 and February 2026 — identified elsewhere in the same filing as OpenAI OpCo, LLC and Meta Platforms, Inc. — each entitling the holder to purchase up to 160 million shares of AMD common stock at $0.01 per share. Combined, that is up to 320 million shares, subject to vesting and exercise conditions.

The structure, as disclosed:

• Warrants vest in tranches as customers reach specified purchase milestones and as AMD’s stock price meets specified targets (including, for some tranches, performance relative to a market index).
• Each tranche becomes exercisable only after specified technical and commercial conditions are met.
• Each warrant expires five years from issuance and is not settleable in cash; exercise means AMD issues shares.
• As of 26 September 2026, AMD reported a warrant liability of $1.2 billion, representing the fair value attributable to qualifying customer revenue recorded to date.
• During the three and nine months ended 26 September 2026, the change in fair value of the warrant liability produced a non-cash charge of about $1.1 billion ($1,097 million in the earnings-release reconciliation, or $0.66 per diluted share). Prepared remarks describe a “$1.1 billion noncash charge in other income and expense, primarily reflecting the increase in our share price and changes in the estimated probability of meeting the warrant conditions.”
• AMD says these quarterly non-cash fair-value adjustments are expected to continue and will vary with stock price and valuation assumptions; they do not affect non-GAAP results or free cash flow.
• The 10-Q states that, as of 26 September 2026, no warrant shares had vested or become exercisable, or been exercised, forfeited or expired. The liability is an accounting mark, not shares already issued; the filing treats the warrant shares as contingently issuable until vesting and exercise conditions are met.

Reuters, reporting the same day as earnings, framed the economics in commercial language AMD’s IR pages do not fully restate in the press release: a February arrangement to sell Meta up to about $60 billion of AI chips over five years alongside warrants that could reach about 10% of the company if milestones are hit, following a similar OpenAI structure. Those commercial headlines are useful context; the primary description of share counts, strike, vesting and the $1.1 billion charge is AMD’s 10-Q and earnings materials. Where Reuters goes beyond the filing (for example, the “$60 billion” Meta headline), treat it as secondary unless AMD restates it.

Contractual versus projected — draw the line

The distinction matters because equity-for-silicon deals are easy to overread. A warrant that can deliver roughly 10% ownership if every milestone clears is not the same as 10% ownership today. A customer “running models on Helios” is not the same as a take-or-pay for the next five years of training clusters. And Q4 guidance of ~$17 billion is a management expectation, not a booked backlog.

Why labs accept the trade

From the lab or hyperscaler side, the bargain is straightforward. Scarcity of leading-edge GPUs has been a strategic risk since 2023. Paying partly in potential equity softens cash outlay per effective watt of new capacity, aligns the supplier with the customer’s deployment success, and creates a second architecture that pricing and allocation negotiations can lean on. AMD, for its part, buys demand visibility and a reference customer list that enterprise and sovereign buyers notice.

The risk is mutual. If purchase milestones slip, warrants do not vest — AMD keeps the upside of its share price without issuing stock, but also misses the volume those milestones implied. If milestones clear and the stock is high, GAAP earnings absorb ongoing fair-value noise and eventual dilution. Customers that become large shareholders also become counterparties with conflicting incentives: they want cheap chips and a valuable AMD stake.

What to watch next

1. Helios volume in Q4 and early 2027. Guidance implies a step-up; the question is how much of Data Center growth is MI450/Helios versus EPYC CPUs.
2. Vesting disclosures. Future 10-Qs should show whether any tranches vest or become exercisable — the moment equity-for-silicon stops being theoretical.
3. Software stickiness. ROCm 10 and Helion matter only if frontier training and inference stacks stay on AMD after the first racks.
4. Peer response. Nvidia’s financing and ecosystem moves, and custom silicon from Google, Amazon and Microsoft, set the ceiling on how large a merchant second source can become.

Bottom line

AMD did not dethrone anyone in a single quarter. What it did, on 3 November 2026, was publish evidence that the second-source strategy has left the slide deck: record Data Center revenue, named frontier and cloud customers on Helios, and a billion-dollar accounting bill for the warrants that helped get them there. Equity for silicon is neither free demand nor guaranteed ownership. It is a contingent contract — and Q3 was the first quarter in which both the silicon and the contingency showed up clearly in the same set of filings.

Related reading

• Capex Arms Race: Mapping Hyperscaler AI Spend and What It Buys
• The AI Accelerator Market: GPUs, ASICs and the Second-Source Race
• The $895 Billion Compute Book: Mapping Anthropic’s Supplier Stack
• Multi-Trillion-Parameter Training: What the Next Silicon Generation Must Deliver

Images: original illustrations generated for AndBeyond.Tech; no logos or real people depicted.