OpenAI is building its own chips now. It’s becoming a bit of a tradition in the AI industry to start yet another week with another 10 gigawatt announcement from OpenAI. Shall we start calling it Ten Gigawatt Tuesday?
Here's what happened: OpenAI just announced a massive partnership with chip giant Broadcom to design and deploy its own custom AI accelerators. The goal? To build what they call the biggest industrial project in human history: a 10-gigawatt AI super-scaffolding to power the next generation of AI and finally start chipping away at the global compute shortage.
- The partnership will deploy 10 gigawatts of custom OpenAI-designed AI accelerators and networking systems, with deployment starting in late 2026 and completing by 2029 (they’re expecting initial silicon back “very soon”).
- The two companies have been quietly collaborating for the past 18 months, designing not just a custom chip but an entire system optimized for OpenAI's specific workloads.
- Plot twist: OpenAI used its own AI models to help design the chip, achieving “massive area reductions” by having AI optimize components that would've taken human engineers another month to get to.
- And because the infinite money glitch go brr, Broadcom’s stock closed up 9.88% by the end of the day.
For the past 18 months, the two companies have been secretly collaborating on a new chip tailored specifically for OpenAI's workloads. Now, they’re taking it to the next level by building the entire system, from the silicon all the way up to the data center racks.
Here's the plan:
- Custom Silicon: OpenAI is designing the chip, and Broadcom is co-developing and manufacturing it. This allows OpenAI to embed its learnings from building models like GPT-5 directly into the hardware.
- Massive Scale: They're deploying an additional 10 gigawatts of power. For context, OpenAI currently runs on about 2 gigawatts. This new buildout is like constructing several small cities dedicated purely to AI.
- Timeline: The first custom-built systems are scheduled to start rolling out in late 2026 and will continue through 2029.
This isn't just about getting more chips; it's about making them smarter. By controlling the entire stack, from the transistors to the software, OpenAI believes it can achieve huge efficiency gains. Sam Altman's vision is simple: wring out more intelligence per watt, which translates to faster, cheaper, and more powerful models for everyone. Greg Brockman says the current 10 gigawatts is just a "drop in the bucket" compared to what's needed to give every person on earth their own AI agent.
One of the coolest parts? OpenAI is using its own AI models to help design the chips, a process that has apparently led to "massive area reductions" and sped up the timeline.
What to expect: This move is a direct response to the insane demand for AI. As Sam Altman put it, "You optimize by 10x and there's 20x more demand." Cheaper, more abundant compute will unlock features that are currently too expensive for wide release, like personal AI agents that work for you 24/7. This partnership signals OpenAI isn't just a model company anymore; it’s becoming an infrastructure powerhouse dead-set on controlling its own destiny.
The announcement, made in a joint podcast featuring OpenAI CEO Sam Altman, President Greg Brockman, and Broadcom’s CEO Hock Tan and Semiconductor President Charlie Kawwas, frames the project in historical terms. "A lot of ways that you would look at the AI infrastructure build-out right now, you would say it's the biggest joint industrial project in human history," Altman stated, setting a grand tone for the ambitious undertaking.
This partnership is the culmination of 18 months of quiet collaboration. Initially focused on designing a single custom chip tailored to OpenAI's unique workloads, the project’s scope has expanded to encompass the entire system—from the silicon wafer to the fully operational data center rack.
The Strategy: Vertical Integration for Unprecedented Efficiency
At the heart of OpenAI’s decision to design its own chips is the principle of vertical integration. For years, the company has relied on off-the-shelf hardware, primarily GPUs from partners like NVIDIA and AMD, to power its research and products. While incredibly powerful, these general-purpose processors are not perfectly optimized for the specific computational patterns of large language models.
By taking control of the entire hardware and software stack, OpenAI aims to achieve a new echelon of performance and efficiency. "We are able to think from like etching the transistors all the way up to the token that comes out," Altman explained. This holistic approach allows OpenAI to embed what it’s learned from creating frontier models like GPT-4 and Sora directly into the hardware, creating a tightly coupled system where the chip, networking, racks, and algorithms are all designed to work in perfect harmony.
The expected result is a dramatic increase in the amount of "intelligence per watt"—a key metric in an industry increasingly constrained by energy consumption. Cheaper, faster, and more capable models are the ultimate prize, which OpenAI believes will unlock a torrent of innovation and wider accessibility. Hock Tan of Broadcom likened the future infrastructure to foundational utilities like the railroad or the internet, envisioning it as "critical infrastructure...for 8 billion people globally."
Why Now? The Insatiable Demand for Compute
The move is a direct response to a lesson OpenAI has learned repeatedly: demand for AI is functionally infinite. "You optimize by 10x and there's 20x more demand," Altman remarked, highlighting the paradoxical cycle where every efficiency gain immediately creates an even larger appetite for AI services.
As a result, OpenAI has launched an all-out war on the global compute shortage, unveiling a 30-gigawatt infrastructure plan that includes a $100B deal with NVIDIA, a massive GPU purchase from AMD, the current project Stargate, and this new initiative with Broadcom.
Here’s how the other partnerships stack up:
- The NVIDIA Deal: A landmark 10-gigawatt partnership with its long-time collaborator where NVIDIA will invest up to $100 billion in OpenAI as the new systems, built on its next-gen Vera Rubin platform, are deployed.
- The AMD Deal: A 6-gigawatt agreement for multiple generations of AMD's Instinct GPUs. To seal the deal, AMD is giving OpenAI warrants for up to 160 million shares of its stock (a clever way to align interests) which will vest as OpenAI hits purchasing and deployment milestones.
- The Stargate deal: OpenAI's $500 billion, 10-gigawatt infrastructure joint venture with Oracle and SoftBank to build five new data center sites across the U.S. The partnership includes a $300 billion agreement with Oracle alone for 4.5 gigawatts over five years (essentially OpenAI building its own hyperscale cloud empire, with the first site in Abilene, Texas already operational and running NVIDIA's GB200 racks).
This is basically a full-stack attack on compute scarcity. By controlling its own custom chip development with Broadcom, OpenAI can "wring out more intelligence per watt," leading to cheaper and more powerful models. Meanwhile, the massive deals with NVIDIA and AMD ensure they have the raw power and a diversified supply chain to scale immediately
The Strategy: Buy, Diversify, and Build
Together, these three deals paint a clear picture of OpenAI’s strategy. It’s a three-pronged attack designed to corner the world's available AI compute and mitigate risk at every level. First, it is buying the best and most powerful off-the-shelf systems from the market leader, NVIDIA. Second, it is de-risking its dependency on a single supplier by bringing on AMD as a formidable second source. And third, it is building its own future with custom, hyper-optimized silicon through its Broadcom partnership.
All three initiatives are set to begin their initial deployments in the second half of 2026, signaling a massive, coordinated infrastructure expansion.
The underlying motivation is the functionally infinite demand for AI. "You optimize by 10x and there's 20x more demand," Altman has remarked. OpenAI currently operates on roughly 2 gigawatts of power. This 26-gigawatt expansion—a more than 13-fold increase—is deemed necessary not just to serve hundreds of millions of users, but to power the increasingly capable models of the future and unlock features like persistent, personal AI agents for everyone on Earth.
By securing its own supply chain on an unprecedented scale, OpenAI is not just building for its users; it's building to unshackle its own researchers and accelerate the journey to AGI. This is a foundational play for the future, an all-in bet that whoever controls the compute, controls the destiny of artificial intelligence.
The Acceleration Question: Why This Compute Matters
OpenAI currently operates on a capacity of just over 2 gigawatts, which powers ChatGPT for its 800 million weekly active users, the Sora video generation service, its API, and all internal research. The additional 10 gigawatts from the Broadcom partnership, combined with other ongoing projects, will bring OpenAI’s total capacity closer to 30 gigawatts in the coming years.
To put their journey so far in perspective:
- OpenAI started with a 2-megawatt cluster (enough power for a mid-size Costco)…
- Scaled to 2 gigawatts this year (enough to serve 800M weekly users)…
- …And will hit close to 30 gigawatts with these recent partnerships (quite literally all of the power of Australia).
Even this staggering number may not be enough. Greg Brockman described the 10-gigawatt plan as a mere "drop in the bucket compared to where we need to go." The long-term vision is one of "compute abundance," a world where the computational power needed for advanced AI is no longer a scarce and fiercely contested resource. This abundance is seen as a prerequisite for deploying truly transformative technologies, such as personalized AI agents that can work continuously for every individual on the planet—a concept Brockman noted would require "10 billion chips."
This scarcity is felt acutely within OpenAI itself. Access to compute is a constant bottleneck, limiting the rollout of new features and pitting internal research teams against each other for allocations. By securing its own supply chain, OpenAI is not just building for its users; it's building to unshackle its own researchers and accelerate the path to Artificial General Intelligence (AGI).
OpenAI’s 30-gigawatt infrastructure plan is about more than just scaling up today's products; it’s about providing the fuel for a potential paradigm shift in the very nature of technological progress. This raises a critical question at the heart of the AI industry: what happens when AI systems become powerful enough to automate their own research and development? This scenario, known as recursive self-improvement or R&D acceleration, could trigger an unprecedented explosion in the rate of technological advancement.
AI Designing AI: A Glimpse into the Future
One of the most fascinating revelations from the announcement is OpenAI’s use of its own AI models in the chip design process. Brockman shared that their models have been applied to optimize components of the new accelerator, resulting in "massive area reductions" on the silicon and a significant acceleration of the development schedule.
This meta-level application of AI—using intelligence to create the very hardware that will generate future intelligence—is a powerful demonstration of a self-improving feedback loop. According to Brockman, the AI can surface optimizations that human engineers might have on their to-do list but would take months longer to explore. This allows the team to move faster and push the boundaries of what’s possible in semiconductor design.
Crucially, METR (an AI safety research org) released recent research on AI R&D suggests that if AI matches human researchers at AI R&D by 2027, we could see 3x acceleration in progress (experts give this a 20% chance).
The pilot study from METR, in collaboration with the Forecasting Research Institute, surveyed a small group of AI forecasting experts and elite "superforecasters" to gauge their beliefs on two key questions: the likelihood of a dramatic acceleration in AI progress, and the potential for that acceleration to cause extreme societal impacts. METR defined a "rapid acceleration" as a three-fold increase in the rate of AI progress. To make this tangible, they framed it as compressing the last three years of AI improvement—the entire journey from a pre-ChatGPT world to the recent release of GPT-5—into a single year.
The study found a consensus that such an event, while not a certainty, is highly plausible. AI experts gave it a median 20% chance of happening by 2029, while the more conservative superforecasters put the odds at 8%. Both groups agreed that the key indicator to watch for isn't just AI's performance on short-term benchmarks, but its ability to successfully conduct open-ended, month-long research projects at or above the level of top human researchers. If AI can truly automate the long, arduous process of scientific discovery, the feedback loop could begin.
Where the two groups diverged sharply was on the consequences of such an acceleration. The study asked them to predict the likelihood of "extreme societal events," such as a doubling of global energy consumption in a single year or an increase in catastrophic risks on the scale of the COVID-19 pandemic.
The results revealed a massive gap in worldview:
- AI experts saw a significant chance of transformative impact, assigning a median probability of over 18% to all extreme outcomes. Their reasoning generally equated a 3x acceleration with the arrival of AGI, an event they believe would fundamentally reshape the world.
- Superforecasters, known for their calibrated and often more skeptical predictions, were far less convinced. They assigned a median risk of less than 0.5% to all extreme events, citing real-world constraints like physical barriers, energy limitations, and human intervention that would likely temper the impact of even a super-intelligent AI.
This stark disagreement highlights the profound uncertainty at the frontier of AI. Even when the experts agree on what technological milestones to watch for, they have vastly different opinions on what those milestones will mean for humanity.
This is the high-stakes context for OpenAI's 30-gigawatt gamble. The company is building the very infrastructure that could enable this R&D acceleration. Whether this leads to a controlled burn of incredible progress or an uncontrolled fire of societal disruption is the central debate of our time. OpenAI's strategy suggests they aren't just aiming to find out; they are building the capacity to steer it.
A New Era of Custom Silicon
OpenAI’s venture into chip design reflects a broader industry trend where major technology companies, from Google (TPUs) to Amazon (Trainium/Inferentia) and Microsoft, are increasingly bringing silicon development in-house to gain a competitive edge. It also represents a frustration born from experience. Brockman recalled the early days around 2017 when OpenAI realized the paramount importance of scale but found it difficult to influence the roadmaps of existing chip startups. "Honestly, a lot of them just didn't listen to us," he said.
By partnering with Broadcom, a company with deep expertise in manufacturing and deploying complex systems at scale, OpenAI can now directly realize its own vision for AI hardware. The collaboration aims to create a platform optimized not just for today's models but for the architectures of tomorrow. The systems will be scaled using standards-based Ethernet, a specialty of Broadcom, ensuring open and efficient connectivity within the massive data centers.
As the world stands on the cusp of another leap in AI capability, this partnership is more than a business deal; it is a foundational play for the future. OpenAI is betting that by controlling its own destiny from the sand to the system, it can build the engine that will power the next industrial revolution.
Why This Matters
Allow us a slight tangent, then we'll land the plane. We’ve really been digging Alex Kantrowitz’s podcast lately, as he presents a very (in his words) “level-headed” discussion around all the craziness happening in the AI world.
In a recent episode around the 10:12 mark, Alex and his guest Ranjan discuss AGI using the definition that big AI labs are adopting: “AI that can automate 50% of white collar work.”
And honestly? Today's AI capabilities, once fully developed and deployed, might actually hit that mark (thus justifying the wild datacenter spending). But here's the catch everyone's missing: AI isn't automating full jobs… it's automating full tasks. And those are wildly different things.
METR also recently released fascinating data showing that Claude Sonnet 4.5 can complete tasks that take human software engineers up to 1 hour and 53 minutes with 50% reliability. For context, GPT-4 could barely handle 10-minute tasks. As we’ve shared before, this capability has been doubling every 7 months for the past six years.
BUT, and this is crucial, tasks aren't jobs. Jobs = responsibility, which = making judgment calls, understanding decisions in a broader context, and being accountable when things go sideways.
AI is becoming the new SaaS, not the new worker. It's a productivity tool that makes humans better at their jobs, not a replacement for the human doing the job (yet). Just like Excel automated calculations but didn't eliminate accountants, today’s AI will automate tasks but won't eliminate knowledge workers. So the question you gotta ask yourself is… is this level of data center build out worth it if it’s only for SaaS 2.0?
Think about it: the reason OpenAI is signing all these deals is to establish its own hyperscale level cloud, and as a byproduct, reduce reliance on NVIDIA, the AI industry's most singular point of failure (controlling 90%+ of the AI chip market), by making AMD and Broadcom legit competitors, and by extension, grow the resilience of the whole industry.
But here’s the problem: because America is basically just one big bet on AI this point, all this attention on AI could inflate the AI bubble even bigger (because, TBH, what else are you going to invest in atm?). This is only a problem if, as Noah Smith writes, any new AI model even just “slightly disappoints” on launch.
The next turn in the launch cycle will be the release of Gemini 3.0, which if this leaked screenshot is to be trusted, will be officially released on Oct 22 (you never know though for sure, though). Let’s hope it does at least sliiightly better than disappoint!
