Amazon Finishes $50bn OpenAI Deal, AWS Locks In $100bn Spend

Amazon has completed its $50 billion investment in OpenAI, disclosing in a quarterly filing with the Securities and Exchange Commission on Friday that it invested $15 billion in the first quarter, a further $13.7 billion in the second, and the remaining $21.3 billion at some point after 30 June. The disbursement closes a structure announced on 27 February, under which $15 billion was committed upfront, and $35 billion was held back against conditions that were never made public.

The completed investment gives Amazon a preferred equity holding in the most valuable private technology company, alongside a commercial agreement that returns much of the capital as contracted revenue. OpenAI has committed to spending approximately $100 billion with Amazon Web Services over eight years, with AWS serving as the exclusive third-party cloud distribution provider for OpenAI Frontier. Andy Jassy, Chief Executive Officer of Amazon, told CNBC when the partnership was announced that the position would “yield a good return for Amazon over a long period of time”.

The trigger that released $35 billion remains undisclosed, and the compressed timeline invites scrutiny

Amazon’s original agreement stated that obligations would terminate if the $35 billion had not been invested by 31 December 2028, subject to acceleration in certain circumstances. The company said at the time that the remainder would follow when OpenAI either went public or met “specified milestones”. OpenAI has not gone public, and Amazon gave no account in the filing of what satisfied the condition.

The Information had reported that one of the milestones could involve OpenAI reaching artificial general intelligence, a report Amazon declined to comment on. Almost three years of optionality were therefore compressed into roughly five months without public explanation, in a transaction large enough to move both companies’ balance sheets. Reports this week that IPO preparations are under way give the timing an additional reading, and shareholders are entitled to ask which of the two paths the money followed.

The equity cheque buys Amazon a contracted customer for its own silicon.

The February agreement covered a full commercial programme. AWS and OpenAI are co-creating a Stateful Runtime Environment powered by OpenAI models, available through Amazon Bedrock, giving developers persistent context, memory and compute across long-running workflows. AWS also became the exclusive third-party cloud distribution provider for OpenAI Frontier, the platform through which organisations build, deploy and manage teams of AI agents.

Underneath the software sits the compute. The two companies expanded an existing $38 billion agreement by $100 billion over eight years, with OpenAI committing to consume approximately 2 gigawatts of Trainium capacity spanning both Trainium3 and the Trainium4 generation expected to begin delivery in 2027. Sam Altman, Co-Founder and Chief Executive Officer of OpenAI, said at the announcement that both companies believe “AI should show up in ways that are practical and genuinely useful” for people. The agreement secures guaranteed capacity for OpenAI at a lower cost of producing inference at scale. It secures for Amazon a demand anchor for custom silicon that spent several years struggling to prove itself against Nvidia in the open market.

Amazon’s second quarter explains the urgency behind a commitment of this size

AWS revenue grew 36.7% year on year to $42.2 billion in the quarter ended 30 June, the fifth consecutive quarter of acceleration, taking the annualised run rate to $169 billion against a contracted backlog of $496 billion. Group revenue reached $200.6 billion, up 20%, with operating income of $27.5 billion, up 43%. Capital expenditure for the quarter totalled $54.21 billion, up 68%, and management raised full-year guidance to $220 billion while expecting capacity constraints to persist into 2027 and 2028.

Jassy described AWS as “booming” on the earnings call and pointed to the artificial intelligence and chips units, each of which passed a $25 billion annualised revenue run rate while growing at triple-digit rates. Anthropic and OpenAI, which he described as the two leading AI labs in the world, have each made multi-year, multi-gigawatt capacity commitments on Trainium. He characterised the adoption curve as a “barbell”, with AI labs at one end and enterprises pursuing cost avoidance at the other, and the largest segment, existing enterprise production workloads, still ahead of the market.

Suppliers funding their largest customers has become the structural pattern of the AI market

Amazon’s $50 billion sat inside a $110 billion round at a $730 billion valuation, alongside $30 billion each from SoftBank and Nvidia. The pattern across all three investors is consistent. Capital moves from suppliers into their largest customer and returns to those suppliers as contracted revenue, chip orders and cloud backlog. Investors assessing AWS backlog growth, Nvidia’s order book and SoftBank’s marks are now assessing overlapping claims on the same underlying demand, and the circularity has spread well beyond a handful of deals.

For enterprise buyers the consequence shows up in procurement. Microsoft, which first backed OpenAI in 2019 and has committed more than $13 billion, retains its exclusive licence and access to intellectual property across OpenAI models and products, so the distribution of OpenAI capability now runs across the two largest cloud providers under different commercial terms. Teams that assumed a single-vendor route to frontier models will need to account for that split, and for the possibility that pricing on agent platforms comes to depend on which hyperscaler holds the underlying capacity commitment.

ChatGPT is meanwhile approaching one billion weekly active users, a milestone reached roughly seven months later than OpenAI had projected. The demand supporting a $730 billion valuation and $100 billion of cloud spending is real and growing, and the eight-year horizon on the AWS agreement requires it to hold for considerably longer than any AI product cycle has run so far.