The OpenAI-AMD deal separates purchases, milestones, and shares

On October 6, 2025, AMD and OpenAI announced an agreement to deploy up to six gigawatts of Instinct processors across several product generations. The first phase—one gigawatt of MI450 systems targeted for the second half of 2026—was tied to a binding purchase commitment. The remainder depends on future purchases and conditions, and must not be presented as hardware already delivered.

The deal added an unusual incentive: AMD issued OpenAI a warrant covering up to 160 million shares. It did not immediately hand over those shares or ownership of nearly ten percent of AMD. A warrant is a contractual right to acquire stock if vesting and exercise rules are satisfied. Understanding the transaction requires keeping purchase, delivery, vesting, exercise, and ownership separate.

The press release describes the plan; the filing bounds the commitment

The joint announcement describes a definitive, multiyear, multigeneration agreement. It calls AMD a core strategic compute partner, targets the second half of 2026 for the first gigawatt, and extends the collaboration from MI450 to future products. It also says the companies will share technical expertise to optimize their hardware and software roadmaps.

The word “definitive” does not make every headline number unconditional. AMD’s Form 8-K filed with the SEC says OpenAI made a binding commitment at signing to purchase the initial gigawatt of MI450 products, directly or through affiliates or authorized purchasers. Full vesting of the warrant requires purchases to reach six gigawatts.

This difference produces a reading rule: every scale needs a verb and a status. One gigawatt was committed; six was the intended cumulative scale and a condition for the maximum incentive; the first delivery was still in the future. “Agreed,” “purchased,” “manufactured,” “delivered,” “installed,” “energized,” and “operational” are not synonyms.

A gigawatt does not reveal how many chips exist

The agreement expresses capacity as power associated with GPUs, not as unit count. Neither the announcement nor the 8-K publishes accelerator quantity, power per device, rack design, location, total facility demand, or annual energy. They also do not support adding six gigawatts to every other announcement without checking overlap, schedule, and measurement boundary.

Accelerator power does not automatically equal facility power. Host servers, memory, networking, storage, electrical conversion, and cooling require additional resources. Energy also requires time and utilization. Follow-up reporting should therefore record what the gigawatt measures and separate nameplate capacity, connected capacity, and workload actually used.

The first physical evidence will be delivery of the initial gigawatt because the 8-K links that event to the first stock tranche. Technical acceptance, commissioning, and live workloads must follow. The announcement does not yet demonstrate performance against another platform, cost per task, or sustained availability.

The warrant is a right, not an ownership stake received

The warrant agreement gives OpenAI the right to purchase up to 160 million AMD common shares at an exercise price of one cent per share, subject to adjustments. The instrument was issued as a material inducement for OpenAI to enter the supply agreement and make substantial product purchases.

There are three distinct states. Potential shares begin unvested. A tranche becomes vested when its milestone schedule is satisfied. It becomes exercisable only after the exercise conditions are also met. Finally, the holder exercises the right and receives shares. Saying OpenAI “obtained ten percent” skips those stages and turns a conditional maximum into present ownership.

The one-cent price does not mean the economic value is transferred without consideration. The warrant is tied to the purchase agreement and commercial and technical milestones. The contract itself says it is intended to encourage product purchases and allows either cash payment or cashless exercise under a formula. GPU spending, compliance with conditions, and potential dilution are all part of the transaction’s economics.

The public conditions contain several gates

The 8-K supplies the general architecture. The first tranche vests after delivery of the first MI450 gigawatt. Later tranches depend on purchases scaling to six gigawatts. AMD stock-price targets, rising to 600 dollars per share for the final tranche, stock-performance thresholds, and other technical and commercial requirements must also be satisfied before exercise.

The detailed vesting and exercise schedules are redacted in Exhibits E and F of the public agreement. It is therefore impossible to reconstruct how many shares correspond to each gigawatt, which technical tests trigger each tranche, or every relevant date. A summary that invents a linear progression would fill withheld terms with assumptions.

The instrument expires at 5 p.m. Eastern time on October 5, 2030. At that point, unvested shares and vested but unexercisable shares are canceled; remaining exercisable shares are deemed exercised under the contract. Early termination of the supply agreement for cause can create similar effects sooner. Expiration limits the time available to complete the sequence.

The percentage needs a denominator and date

One hundred and sixty million shares can be compared with AMD shares outstanding to produce an approximation near ten percent, but it is not a fixed percentage. The denominator changes through issuance, repurchases, stock compensation, and other instruments. The warrant also provides adjustments for specified corporate actions.

Verifiable reporting should state share count, denominator, date, and method: percentage of existing shares, fully diluted basis, or post-exercise total. Without all four, “ten percent” sounds like stable ownership. Before exercise it remains potential exposure, not voting rights or title to that block.

If new shares are issued upon exercise, existing owners can be diluted: their fraction of the total falls even if they retain their shares. The actual magnitude will depend on how many warrant shares vest and are exercised, AMD’s capitalization at that time, and any adjustment. The contractual maximum is not a forecast of full issuance.

The revenue outlook is not a published price

AMD said it expected tens of billions of dollars in revenue over the collaboration. This is a company forecast accompanied by risk warnings, not the disclosed value of the initial commitment or an invoice already collected. The full product purchase agreement is not public, and the 8-K gives no system price, discounts, payment schedule, or expected margin.

Revenue would be recognized according to deliveries and accounting rules, not merely by announcing total scale. Evaluating the forecast will require following orders, inventory, segment sales, customer concentration, and audited commentary without automatically assigning all AMD growth to OpenAI. Profitability also depends on manufacturing, integration, and support costs.

The warrant and product purchases are linked, but they cannot be netted through improvised arithmetic. Warrant value changes with the share price, milestone probability, and exercise timing. Product pricing remains confidential; estimating the warrant does not reveal what OpenAI pays for compute.

A milestone board makes the deal testable

The first column should contain the obligation: initial purchase, additional purchases, delivery, technical condition, commercial condition, and stock-price target. The second identifies the responsible party. Then come target date, documentary evidence, quantity, status, and effect—vesting, exercise, or revenue. A row without evidence remains “announced,” not “executed.”

For hardware, future evidence is order, fabrication, delivery, installation, power, acceptance, and use. For the warrant, it is milestone certificates, vested share count, satisfaction of conditions, exercise, and issuance. For the business, it is recognized revenue and concentration. Each lane answers a different question and prevents good news in one from validating all the others.

The transferable skill is reading a hybrid deal as several connected contractual machines. First identify which purchase is binding; then which delivery activates each right; next which conditions make it exercisable; finally measure ownership, dilution, and real operation. The OpenAI-AMD deal may reach enormous scale, but its truth will emerge milestone by milestone—not by pretending its maximum has already happened.