The Reported Stargate Supercomputer and AI’s Compute Bottleneck

Sam Altman has suggested that compute could become a currency of the future. I think the idea is worth considering if AI comes to play a larger role in the economy. In April 2024, I am examining a report about a proposed Microsoft–OpenAI data-center project called Stargate, and why its scale matters.

Reports have also described Altman exploring extremely ambitious funding for chip manufacturing, with figures as high as $7 trillion discussed. That is a reported fundraising ambition, not money known to have been raised. I interpret the focus on chips as a response to the cost and availability of the hardware needed to train and run models.

I look at Nvidia’s stock-price rise through early 2024 and attribute much of the excitement to demand for AI chips. Historical market prices and rankings change, and the growth has multiple drivers. My concern for model companies is that buying scarce infrastructure can pressure margins as competition increases.

I divide the AI industry into four broad layers: chips, infrastructure, models and applications. Chips are the processors; infrastructure includes monitoring and other developer tools; models supply capabilities; and applications put them in the hands of businesses and consumers. These categories overlap, but they help explain where I see value accumulating.

In my view, competition from open-weight and proprietary models may make the model layer more commoditized, while chips capture a lot of value. I point to the models available at the time as examples. That is my commercial interpretation, not proof that all models are interchangeable or that model providers cannot create lasting value.

The Information [reported in March 2024](https://www.theinformation.com/articles/microsoft-and-openai-plot-100-billion-stargate-ai-supercomputer/) that Microsoft and OpenAI were considering a data-center project that could cost as much as $100 billion, including a supercomputer called Stargate with a possible 2028 launch. The report described preliminary estimates and private discussions; it was not a company announcement or a confirmed construction budget.

The report said Microsoft would likely finance the project and described Stargate as the fifth phase of a broader plan. A smaller fourth-phase system was reportedly targeted around 2026, with the companies then in phase three. These were reported intentions, not evidence that the systems or financing were finalized.

I see the proposal as part of an infrastructure race and think Microsoft is investing heavily to avoid missing the next technology wave. I also question the companies’ relationship and oversight. My statement in the video that Microsoft owned 49% of OpenAI should not be treated as a verified description of control or of their complex partnership.

For AI infrastructure, I keep returning to two inputs: chips and electricity. I show an illustration of large data centers alongside nuclear power plants as one possible future arrangement. I speculate about grid pressure and security benefits from distributed facilities, but those outcomes depend on the site and grid; the illustration is not evidence of an approved project or a guarantee against attacks.

Nvidia was a major chip supplier in the market I am describing, but I also discuss potential competitors such as Groq and its inference-focused LPU architecture. My claim that GPUs cannot match its speed is a generalization: comparisons depend on the model, hardware, workload and measurement.

In an interview excerpt, Altman argues that compute demand is unlike a fixed market for replacement phones. At a lower price, he expects people to use more intelligence for tasks such as processing email as well as research; at a higher price, they might reserve it for fewer purposes. He identifies energy, data-center construction, supply chains and chip fabrication as challenges.

I connect that argument to the need for more energy capacity, then play an Elon Musk interview excerpt. Musk predicts that electrical step-down transformers will become a bottleneck after AI chips, followed by electricity shortages. His timeline is a forecast rather than a demonstrated result; electrical transformers and AI transformer models are different things despite sharing a name.

I ask how someone could respond if compute becomes more valuable. I mention investing in companies and wonder whether a personal GPU might appreciate, but I do not know the answer. These comments are speculation, not evidence of future returns.

Other companies were also pursuing compute capacity. I point to Meta’s 2024 plan to acquire 350,000 Nvidia H100 GPUs and argue that access to powerful systems may help attract AI researchers as well as support model development. Planned purchases do not by themselves establish delivery or utilization.

Google also develops its own TPUs, and Groq founder Jonathan Ross previously worked on Google’s TPU effort. I end by worrying that a race among already large companies could concentrate economic power. The outcome is uncertain, but access to chips, energy and talent is central to the competition I see developing.