In January, when OpenAI agreed to spend $10 billion on computing capacity from chip startup Cerebras Systems, industry observers called it a bold but calculated move to diversify away from NVIDIA. Four months later, it has doubled down.
According to a report by The Information, OpenAI has committed to spending more than $20 billion over the next three years on servers powered by Cerebras chips. The total outlay could rise to $30 billion, and may grant OpenAI up to a 10% stake in Cerebras.
In terms of AI infrastructure funding, the deal size remains modest—far below the eye-gauging $500 billion Stargate Project or Meta’s $600 billion AI infrastructure expansion. But it could become a defining AI symbiosis, transforming the relationship from supplier and customer into something far more strategically entangled.
Inside the Deal
The structure of the arrangement is notable for its complexity.
Under the terms reported, OpenAI will not merely purchase chips; it will also provide Cerebras with approximately $1 billion to fund the construction of new data centres that will run OpenAI’s own AI workloads. In effect, OpenAI is financing the very infrastructure it will later pay to use.
It has also committed to purchasing 750 MW of compute capacity from Cerebras to run its AI inference workloads.
Cerebras is known for its Wafer-Scale Engine chips, processors that are physically 58 times the size of NVIDIA’s high-end Blackwell GPUs and promise significantly faster inference at lower power consumption. The company swung from a $482 million net loss in 2024 to a $238 million profit in 2025, with revenue climbing 76% to $510 million.
Its valuation also tripled between September 2025 and February 2026 to $23 billion.
On the same day the deal was reported, Cerebras filed for an initial public offering, targeting a valuation of $35 billion and planning to raise $3 billion. OpenAI CEO Sam Altman is an early personal investor in the company.
Inference is the New Battleground
The OpenAI–Cerebras deal reflects a fundamental shift. For years, the race was about training, feeding vast datasets into ever-larger models using NVIDIA’s GPUs. That era is giving way to a new one: inference, the process by which models actually generate responses for users in real time.
Deloitte expects it to account for two-thirds of all AI compute power by the end of 2026. Unlike training, which runs in controlled batches, inference must be instant, reliable, and massively scalable.
“There will be billions of dollars’ worth of specialised chips optimised for inference, but they’ll sit in data centres and enterprise servers as well, and some will use as much or even more power than general-purpose AI chips do,” Deloitte noted in its TMT Predictions 2026 report.
But inference costs have also declined, with Gartner forecasting a 90% drop by 2030 compared to 2025 levels.
“The network is the nervous system in an AI-driven enterprise. It determines how fast you can move, how much you spend, and whether your AI investments produce value,” Kate Johnson, Chief Executive of Lumen Technologies, a global fibre-network and technology services provider, tells AIM.
The Infrastructure Arms Race
The wider tech industry is in the midst of an infrastructure spending spree without historical precedent. According to analysis by Futurum Group, the five largest US cloud and AI providers—Amazon, Alphabet, Microsoft, Meta, and Oracle—have collectively committed to spending between $660 billion and $690 billion on capital expenditure in 2026 alone, nearly doubling 2025 levels.
The World Economic Forum’s 2026 CEO Survey underlines why boardrooms feel compelled to act at this scale: “The scale of investment in AI infrastructure is already reshaping competitive dynamics.” Those who delay, the survey suggests, risk falling too far behind to recover.
“The question every CEO should be asking is not ‘how can we use AI?’ but rather ‘do we have the infrastructure to use AI at scale?’ That is where the real competitive battle will be won or lost. For the past few years, the AI conversation has been dominated by models—but 2026 marks a turning point: the companies winning with AI are not necessarily those with the best models, but those with the best infrastructure to deploy, manage, and scale them,” comments Neeraj Kumar, Co-founder and Chief Revenue Officer of investment advisory firm Superhuman Ventures.
Models are becoming commoditised. The performance gap between leading models has narrowed dramatically, and open-source alternatives are catching up faster than anyone predicted.
“Meanwhile, the infrastructure to operationalise these models—orchestration layers, data pipelines, monitoring systems, and deployment frameworks—remains a significant differentiator. Companies that master the infrastructure layer will be able to leverage any model that emerges, making them anti-fragile to the rapid pace of AI innovation,” he adds
For OpenAI, specifically, the Cerebras deal serves a dual strategic purpose. The first is simply securing access to the computing capacity it needs to keep pace with demand. The second is more structural: reducing its dependence on NVIDIA, which remains the dominant supplier of AI chips but whose products are in perpetually short supply and command premium prices.
What distinguishes the arrangement from a conventional procurement deal is the depth of financial entanglement. OpenAI is simultaneously a customer, an investor, and a lender to Cerebras. Its spending commitments are the central pillar of Cerebras’s IPO narrative.
The pattern is not unique to this deal. Across the industry, leading AI companies are moving away from transactional chip procurement and towards long-term, equity-linked infrastructure partnerships—securing the compute they need while staking financial claims in the suppliers that provide it. Amazon has committed $25 billion to Anthropic. Google has directed more than a gigawatt of TPU compute to the same company. The infrastructure layer, once an afterthought in the AI conversation, has become the newest competitive terrain.
“Organisations that invest only in hardware and models—and neglect the organisational capability to use them—will find that their infrastructure becomes a liability rather than an asset. The next phase of competitive advantage belongs to those who balance automation with augmentation, and who treat infrastructure investment as inseparable from workforce investment,” said Devendra Jain, Lead, AI and Frontier Technologies at World Economic Forum, during the AI & Business Innovation Summit in London last month.
In the infrastructure arms race that now defines the AI industry, speed, capital, and control of the supply chain have become inseparable. OpenAI and Cerebras, for better or worse, are now bound together by all three.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.