The Dilemma of Technological Returns in the AI Compute Investment Frenzy

OpenAI updated its compute expenditure expectation through 2030 from approximately $600 billion up to $750 billion, while announcing the launch of "Project Camellia," a $20 billion self-operated data center project in Georgia, and signing a 3.2 gigawatt power supply agreement with a local utility company. The expenditure scale of this single enterprise already exceeds the annual GDP of many medium-sized countries. Yet, looking at the historical patterns of technological evolution, the marginal return rate of hardware investment exhibits a clear diminishing trend in the domain of compute infrastructure. Early GPU cluster builds achieved quantum leaps in model performance, but when computing scale expands to tens or hundreds of thousands of chips, engineering constraints like communication latency, heat dissipation efficiency, and fault recovery time systematically erode the effective utilization of theoretical compute power. OpenAI raising its budget by 25% to $750 billion offers no supporting data regarding improvements in unit compute output efficiency. This reliance on simple linear scale expansion has too many precedents in technological history—from the termination of the Superconducting Super Collider to the breakdown of Moore's Law in genomics—where massive investment does not naturally equate to technological breakthroughs.

The 3.2 GW power supply agreement presents another dimension of technological constraint worth scrutinizing in depth. A capacity of 3.2 GW is equivalent to the installed capacity of three nuclear reactors, and the power requirement of data centers is not merely about total volume, but also continuous, stable baseload power supply with an extremely low probability of outages. Whether the power grid structure in Georgia possesses the redundant capacity to handle such a concentrated load, how the massive water consumption from cooling large-scale operational GPU clusters affects local water resources, and whether life-cycle carbon emissions can comply with increasingly strict environmental regulations—these hard engineering constraints will directly determine the actual commissioning schedule and operating costs of "Project Camellia." A common systemic misjudgment in the tech industry is treating infrastructure elements like electricity, land, water, and network bandwidth as infinitely supplied external variables. The reality is that in many regions, the approval cycles and construction timelines for these resources far exceed the iteration cycle of chips; by the time a data center is finally powered on, the GPU models deployed inside may already have been replaced by the next architecture.

The agreement signed between Anthropic and AMD reveals another layer of technological dilemma. Anthropic will procure up to 2GW worth of AMD MI450 chips starting in the first half of 2027, while AMD will invest up to $5 billion in Anthropic. On the surface, this arrangement establishes a deep tie between a chip supplier and a large model developer. However, from a technical compatibility perspective, Anthropic has previously relied heavily on NVIDIA's CUDA ecosystem for model training; migrating its millions of lines of code to AMD's ROCm platform requires extensive operator rewriting and performance tuning. A procurement volume of 2GW means Anthropic must fully migrate all core training tasks to the AMD platform, inevitably incurring the maintenance cost of running two parallel technology stacks during the transition period. More critically, whether the actual compute density and interconnect bandwidth of the MI450 can match the architectural demands of Anthropic's next-generation models cannot be verified until chip tape-out is complete, whereas the procurement volume agreed in the contract has already locked in capital commitments for several years—this time mismatch leaves technical decisions frozen in advance by financial terms.

Microsoft's expanded cooperation with Mistral AI also merits examination from the perspective of technical effectiveness. Microsoft is providing billions of dollars to support Mistral in building GPU data centers in Europe. However, looking at the technical workflow of model training, the physical location of a compute cluster has no bearing on model performance itself; the political symbolic significance and local data compliance value of European data centers far outweigh their technical contribution. Furthermore, as a representative of the open-source route, Mistral's technical advantage lies in the compactness of its model architecture and inference efficiency. Microsoft pushing it to massively expand GPU clusters essentially forces a company known for software algorithms into a heavy-asset operating model, creating a fundamental mismatch with its core competitiveness trajectory.

News of SpaceX inspecting sites for large data centers in Texas extends this compute arms race into a field previously considered unrelated to AI. SpaceX's core technical expertise lies in aerospace engineering and Starlink communications. Its entry into data center construction follows one of two technical logics: either providing compute nodes for Starlink ground gateways, or attempting to apply thermal and energy management experience from aerospace to data centers. The former logic lacks the demand volume to justify a "large data center," while the latter overlooks the fundamental difference between the extremely high reliability required in aerospace engineering and the cost economics of data centers—aerospace-grade thermal management solutions cost dozens of times more per watt than commercial solutions, making it difficult to satisfy the financial model of a commercial data center.

When media analysis highlights that by 2027 the combined capital expenditure of tech giants like Microsoft, Google, Amazon, and Meta will exceed their free cash flow, the technical implications of this signal are far more severe than the financial ones. Capital expenditure exceeding cash flow means companies must continuously rely on external financing to maintain their pace of compute expansion, while investor confidence rests on the dual narrative of continuous leaps in model capability and exponential growth in commercial revenue. Once the performance gain of next-generation foundation models falls short of expectations, or enterprise willingness to pay for AI features reaches a temporary saturation point, a situation of oversupplied compute will quickly form. At that point, massive depreciation expenses will eat into tech R&D budgets, creating a vicious cycle where hardware investment squeezes out software innovation. Even more fundamentally contradictory is that all current compute investments are built on the premise that the Transformer architecture can continue to scale. Yet, every architectural revolution in the history of deep learning—from convolutional neural networks to attention mechanisms—has been accompanied by the partial write-off of massive hardware investments from the previous technological route. The current $750 billion compute expenditure is betting that the entire technological paradigm will not undergo a fundamental change over the next five years, which constitutes the largest systemic technical risk in a field like artificial intelligence where disruption is the norm.