- OpenAI signs a $38.000 billion, seven-year deal with AWS to secure massive computing with Nvidia GB200/GB300 and EC2 UltraServers.
- Full deployment before 2026 and option for later expansion; dedicated capacity already in use and new facilities tailored to your needs.
- Key diversification move beyond Microsoft, against a backdrop of $1,4 trillion in infrastructure commitments and signs of a possible bubble.
OpenAI has made a splash in the tech world by closing a multi-billion dollar deal with Amazon Web Services to secure long-term computing power. We're talking about an agreement valued at 38.000 million which guarantees immediate and growing access to the AWS cloud to run and scale the most demanding artificial intelligence workloads of the company led by Sam Altman.
Beyond the figure, what's relevant is the scope: OpenAI will be able to leverage AWS's world-class infrastructure, with full deployment planned before the end of 2026 and the option to expand starting in 2027. The agreement represents a diversification of providers compared to its historical dependence on Microsoft, while also It strengthens training and inference skills. from ChatGPT and the upcoming frontier models that the company has in development.
What exactly have OpenAI and AWS signed?
The agreement is multi-year, with a seven-year horizon, and stipulates that OpenAI will begin using dedicated AWS capacity immediately. According to both companies, part of that separate capacity is already operational and the rest will be activated gradually until completion before the end of 2026, with room to expand in subsequent years depending on demand.
Operationally, OpenAI will rely on existing AWS data centers, and in parallel, Amazon will build facilities specifically designed for its needs and geared towards cloud nativeThese new locations will incorporate a sophisticated architectural design designed to maximize performance and energy efficiency, a critical aspect when the required computing power grows at a breakneck pace.
The alliance includes the use of Amazon EC2 UltraServers, an architecture that groups large clusters of GPUs within the same network to minimize latency between nodes. In practice, this means that the chips of Latest generation Nvidia (GB200 and GB300) They will be able to communicate with minimal delay, which is key to training giant models and serving inferences on a global scale without bottlenecks.
The stated objective of the parties is clear: to provide OpenAI with a massive, reliable and secure computing base to support training of new frontier models, support ChatGPT traffic, and deploy more complex workloads, including agentive ones, with the elasticity that only a hyperscale like AWS can provide.
- Duration and value: $38.000 billion during seven years, with full deployment before 2026 and option for subsequent growth.
- Dedicated capacity: part available from day one and the rest on ramp, supported by current data centers and in new, custom-built facilities.
- Architecture: Amazon EC2 UltraServers with Nvidia GB200/GB300 GPUs and clusters on the same network for ultra-low latencies.
- Intended uses: training of next-generation models, ChatGPT inference and high-density AI workloads.
The technical dimension: chips, clusters and performance
Training language models and multimodal systems of the size that OpenAI handles requires grouping hundreds of thousands of GPUs y high-performance clusters and connect them to a very high-speed network. AWS's proposal fits this pattern: it integrates Nvidia GB200 and GB300 accelerators into UltraServers connected within the same mesh, in order to reduce communication latency between chips and accelerate both distributed training and massive response serving.
When working at massive scale, every millisecond shaved off internal communications can translate into weeks of training time saved. That's why AWS emphasizes interconnections, queue management, and resource orchestration to ensure that efficiency per chip and per cluster be as high as possible, avoiding wasted capacity and ensuring sustained use.
The design also takes into account mixed scenarios: not only state-of-the-art training, but high-performance inference For production products like ChatGPT, as the user base grows and features become more complex, cost per request and perceived latency for the end user become essential metrics for maintaining competitiveness.
In this context, AWS boasts a track record of operating massive clusters in sectors such as biotechnology, climate research, and finance. That experience—in projects that require more than half a million chips in coordinated operation—it is now being transferred to a client that demands a level of stability and security unusual even by enterprise cloud standards.
The result is a platform designed to scale rapidly, capable of absorbing peak demands and sustaining continuous training and inference loads for months. For OpenAI, having elasticity and availability This caliber is not a luxury, but a prerequisite to continue pushing the bar for cutting-edge AI without slowing down its roadmap.
Strategic movement and relationship with other cloud providers
From 2019 to 2023, OpenAI outsourced all its computing to Microsoft, its largest investor. That agreement included a right of first refusal whereby I could only turn to other providers Cloud services with Redmond's blessing. This constraint has recently been lifted after renegotiating the terms, so OpenAI can freely sign with any cloud platform, a decision that opens the door to agreements like the one we are discussing with AWS.
In parallel, OpenAI has sealed massive purchase commitments with several players in the value chain: an additional agreement has been revealed with Azure valued at $250.000 billionAnother deal with Oracle for $300.000 billion for data centers, and contracts with manufacturers and suppliers such as Nvidia, AMD, Broadcom or Google, in addition to a $22.400 billion agreement with CoreWeave, one of the leading companies in the so-called neo-clouds geared towards AI developers.
Until this announcement, AWS had been the odd one out among the major US cloud providers supporting OpenAI; now it becomes a top-tier partner. The move is significant, because Amazon maintains very close ties with AnthropicAmazon, a direct competitor of OpenAI, has invested billions in the company and is building an $11.000 billion campus in Indiana for its cargo, with large deployments of Trainium2 already in operation.
OpenAI, for its part, continues to prioritize Nvidia accelerators for its most critical workloads, instead of AWS Trainium processors, a trend the market had already been indicating. On the commercial front, the collaboration with Amazon is also evident in the services ecosystem: Amazon Bedrock It integrates multi-house models and, according to the companies, AWS customers can access OpenAI models in that environment for enterprise use cases.
The industry consensus is that diversifying suppliers is not a whim, but a measure of resilience. When computing demand is enormous and sustained, relying on a single cloud This poses an operational and financial risk that should be mitigated. AWS's involvement offers OpenAI redundancy, better negotiation of price and delivery times, and guaranteed continuity in the face of potential chip supply disruptions.
Financial impact, market and implications for the industry
The agreement with AWS is part of OpenAI's unprecedented effort to ensure computing and power at scale. The company has acknowledged commitments worth 1,4 trillions of dollars in infrastructure to build and feed their AI models. That figure, well above the usual standards of the sector, has raised some alarms about the possibility of an investment bubble around the AI economy.
To put it in perspective, Altman has estimated that fulfilling those commitments will require on the order of 30 gigawatts of electrical powerMore than 2% of all installed US capacity by the end of 2023, according to official data. The process of building data centers, securing chip supplies, and closing long-term energy contracts has become the real race behind the AI race.
On the corporate front, OpenAI is completing its metamorphosis from a non-profit structure to a scheme capable of generate profits for investorsThis organizational change paves the way for a possible initial public offering (IPO), a move that various media outlets have estimated at valuations of up to one trillion dollars, always subject to market conditions.
The most immediate financial metrics also have implications. OpenAI is expected to generate around 13.000 million This year, although the company does not expect to reach profitability until 2029. It is a long time horizon which, together with the level of committed capex, explains why some investors are adopting a cautious stance despite the general enthusiasm for generative AI.
The stock market reacted optimistically to the announcement: Amazon shares reached advance at around 5% In the early stages of the session, shares surged, reaching peaks of up to 6% at times, while Nvidia saw gains of around 2,7%, reflecting its role as the dominant provider of AI accelerators. For AWS, the deal represents a validation of its ability to build and operate massive data center networks in the new era of artificial intelligence.
In this context, major technology companies have intensified their capital spending. Last year alone, Amazon, Google, Meta, and Microsoft combined investments exceeding 360.000 million in capex, which fuels the debate about the sustainability of the investment cycle. The circular relationships forged between cloud providers, chip manufacturers, and AI developers—with cross-commitments and long-term acquisitions—are both a growth engine and a source of systemic risk if demand doesn't keep pace.
For AWS, the agreement with OpenAI also has a competitive interpretation: it consolidates its catalog as a reference infrastructure for large-scale AI workloads and, consequently, strengthens its offer For enterprise customers who consume models through managed services. For OpenAI, the gain is twofold: capacity and resilience, with the flexibility to scale without depending on a single provider.
Key points for CTOs and architecture teams
The first lesson is that there is no AI strategy without one computing strategyIf industry leaders are securing clusters with next-generation accelerators and low-latency networks, organizations should follow suit. Prioritize efficiency of their workloads (from training to inference) and take care of metrics such as cost per token and p95 latency as if they were business indicators.
- End-to-end optimization: from preprocessing to training and to serving of inferencesorchestrating resources to reduce waiting times and improve utilization.
- Fault-prepared architectures: design with redundancy, multi-AZ/multi-region and Selective multi-cloud when the risk justifies it.
- Energy and capacity plan: size growth with realistic scenarios, anticipate peaks and adjust commitments flexible purchasing.
- Governance and security: protecting models, data, and pipelines with identity controls, encryption, and continuous audit in hybrid environments.
In operational summary, the move by OpenAI and AWS shows that the AI bottleneck is no longer talent or software, but sustained access to high performance hardwareEnergy and well-orchestrated networks. Whoever achieves these three elements will be able to accelerate product cycles and differentiate themselves.
The companies' public statements all point in the same direction: there is a unprecedented demand Computing power is essential, and to scale cutting-edge models, massive and reliable computing is needed. At the same time, Microsoft maintains a leading role as a major OpenAI partner on Azure, Oracle is expanding its data center footprint, Google Cloud is already among those powering ChatGPT, and CoreWeave is serving AI developer workloads with a tens of billions of dollars in deals.
It remains to be seen how the workload is distributed between platforms and how the actual cost of training and inference evolves as new generations are joining Chips are being optimized and software stacks are being refined. But the trend is clear: diversifying suppliers and building long-term computing reserves are becoming strategic levers to sustain the pace of innovation.
This alliance between AWS and OpenAI not only secures capacity for the coming years; it also confirms that the next phase of AI will be decided at the foundation level: chips, data centers and energycoordinated with architectures that make the most of every millisecond. In this arena, securing agreements of this magnitude is as much a matter of technology as it is of corporate strategy.
