In partnership with
AI Spotlight · The Compute War
The Compute War: Why Big Tech is Betting Everything on Infrastructure
The AI race has a new front. It is not a model. It is not a product. It is a data center, and whoever fills it fastest may decide which AI tools exist in five years.
For the past three years, the AI conversation has been dominated by models. Which one is smarter. Which one writes better. Which one scores higher on benchmarks. But quietly, in the background, a different and arguably more important competition has been accelerating. The race for compute.
In 2026, the four largest technology companies are on track to spend hundreds of billions combined on AI infrastructure, a jump of more than 70 percent from the year before. OpenAI signed a massive multi-year deal with AWS, then expanded it significantly months later. Amazon committed tens of billions directly into OpenAI. Wall Street analysts now project total Big Tech AI capital expenditure could exceed one trillion by 2027. Nvidia's CEO has estimated that multiple trillions will be spent on AI infrastructure by the end of the decade.
This issue explains what is actually happening, why it matters, and what "winning the compute race" means for the AI tools that solo operators and knowledge workers depend on every day.
|
What is the compute war, exactly? Compute refers to the raw processing power required to train and run AI models: chips, servers, data centers, and the electricity to power them. In the current AI landscape, more compute means faster model development, lower inference costs, and the ability to run larger, more capable models at scale. The company that controls the most compute infrastructure does not just win the race today. It shapes which AI products are even possible tomorrow. |
* * *
The numbers
Hundreds of billions in 2026 alone. What that actually means
The combined AI infrastructure spending of Alphabet, Amazon, Microsoft, and Meta in 2026 is projected to land in the mid-hundreds of billions range. To put that in context: it exceeds the entire GDP of Switzerland, Sweden, and Norway combined. Each company is committing record capital to chips, data centers, and energy infrastructure, with year-on-year increases ranging from 8 to 24 percent across the group.
Global AI capital expenditure forecasts from major banks project spending growing at a 25 percent compound annual rate through 2030. The baseline projection from Goldman Sachs places cumulative AI infrastructure investment at multiple trillions between now and 2031. These are not conservative estimates. They represent a consensus view that AI infrastructure is the defining capital cycle of the decade.
|
2026 AI infrastructure spend at a glance Amazon: The largest single-company AI infrastructure commitment in 2026, with record capex and a direct equity investment into OpenAI on top. Alphabet: Heavy investment in its own custom AI chips (TPUs) alongside Nvidia infrastructure, with the second-largest capex commitment of the group. Microsoft: The sharpest year-on-year rise at 24 percent, largely driven by Azure AI capacity expansion and its ongoing relationship with OpenAI. OpenAI: Massive hardware and cloud infrastructure agreements committed over the next decade, including a landmark multi-year partnership with AWS that became one of the largest cloud compute deals in history. |
These are not typical capital expenditure cycles. This is a deliberate, coordinated bet that whoever builds the most compute infrastructure today will control the AI market for the next decade.
* * *
The deal that changed everything
Why OpenAI landing on AWS was the signal nobody expected
OpenAI was built on Microsoft Azure. That relationship was so deep and so public that most of the industry assumed it was permanent. Then in November 2025, OpenAI signed a landmark multi-year deal with AWS to run its core AI workloads on Amazon's infrastructure, giving it access to hundreds of thousands of Nvidia GPUs across multiple US sites already live and ramping fast.
Three months later, in February 2026, the partnership was dramatically expanded. Amazon made a major direct investment into OpenAI. The compute deal grew substantially in scope and duration. OpenAI committed to consuming a significant portion of Amazon's custom Trainium chip capacity. AWS became the exclusive third-party cloud distribution provider for OpenAI Frontier, the platform that lets enterprises deploy teams of AI agents. A single deal became the backbone of the AI infrastructure stack that most of the world runs on.
|
What the OpenAI-AWS deal actually unlocks For OpenAI: Long-term compute capacity without depending on a single partner, lower cost per unit of intelligence at scale, and purpose-built silicon through Trainium alongside Nvidia GPUs. For AWS: The most strategically valuable workload in the AI industry locked into Amazon's infrastructure for nearly a decade, plus new enterprise distribution through OpenAI Frontier on Amazon Bedrock. For the industry: A clear signal that compute infrastructure is now the strategic moat, not the model itself. The model can be replaced. The infrastructure contracts last nearly a decade. |
The deal was also a warning shot to every AI company still operating without long-term infrastructure commitments. If OpenAI, the most capital-backed AI company in the world, needed to lock in a massive compute relationship, what does that say about the floor of investment required to stay competitive?
Where to Invest $100,000 Right Now, According to Experts
Investors face a dilemma. When the S&P 500 finished its worst quarter since 2022 last month, diversifiers like bonds and bitcoin fell too.
Even with the turnaround in mid-April, analysts at Goldman Sachs and Vanguard have projected low-single-digit annualized returns from 2024-2034.
Bloomberg asked where experts would personally invest $100,000 for their March monthly edition.
One answer that surfaced for a second time? Art.
It's what billionaires like Bezos and the Rockefellers have privately used to diversify for decades.
Why?
Appreciation. The ArtPrice100 Index outpaced the S&P 500 overall from 2000 to 2025
Low-correlation. The postwar contemporary segment has moved independently of traditional investments like stocks since ‘95.*
Resilience. A scarce, physical, and global asset class with decades of demonstrated demand.
Thanks to the world's premier art investing platform, now anyone can invest in works featuring legends like Banksy, Basquiat, and Picasso, without needing millions.
Shares in new offerings can sell quickly but...
*According to Masterworks data. Investing involves risk. Past performance is not indicative of future returns. See important Reg A disclosures at masterworks.com/cd.
What winning looks like
What "winning the compute race" actually means in practice
The phrase "winning the compute race" gets used loosely, but it has a concrete meaning. The company that wins controls the cost of producing AI output. More compute means lower cost per token, per query, per agent action. Lower cost means more aggressive pricing. More aggressive pricing means broader adoption. Broader adoption means more data and more revenue to reinvest in more compute. It is a flywheel, and it starts with raw infrastructure.
Wall Street analysts have put it plainly: the race to provide AI compute is the next winner-take-all or winner-takes-most market. That framing matters because it explains why companies are spending at a loss today. They are not trying to profit from infrastructure in 2026. They are trying to lock in the cost structure and capacity that makes every AI product built on top of them cheaper and faster than alternatives built elsewhere.
|
Three things compute dominance actually buys Pricing power: When your cost to serve a query is a fraction of a competitor's, you can undercut on price while still generating margin. Infrastructure cost is the hidden driver behind every headline about cheaper AI models. Model velocity: More compute lets you run more training experiments faster. The companies with the most infrastructure can iterate through model generations at a pace that resource-constrained competitors simply cannot match. Agentic scale: Running a single query is cheap. Running an autonomous AI agent that makes thousands of sequential decisions is not. The companies that can serve agentic workloads at enterprise scale at acceptable cost will dominate the next product cycle. |
* * *
The survival question
Which AI tools survive long-term when compute is this expensive
The compute war creates a brutal sorting mechanism for the AI tool ecosystem. Products that sit at the top of cheap, scalable infrastructure survive and grow. Products that sit on expensive, fragile, or insufficient compute face a ceiling: they cannot improve fast enough, cannot lower prices aggressively enough, and eventually get displaced by tools that can.
The blunt reality is that most independent AI startups cannot compete at this level. They can build excellent products on top of foundation models, but they cannot build the infrastructure layer themselves. The question for every AI company in 2026 is whether their product is differentiated enough to justify the margin compression that comes from depending on Big Tech infrastructure, and whether they can lock in favorable compute relationships before the prices rise.
|
The three tiers of AI survival in a compute-defined market Tier 1: Infrastructure owners Amazon, Google, Microsoft, and Meta. These companies are building the physical layer. They win regardless of which AI model wins, because every model runs on their chips and in their data centers. Their products may not always be the smartest, but their cost structure is unbeatable from the outside. Tier 2: Infrastructure-locked foundation models OpenAI, Anthropic, and ByteDance (Doubao). These companies do not own the infrastructure but have locked in long-term compute relationships at favorable terms and scale. They can compete on model quality and pricing because their cost per unit of intelligence is protected by contract. Their survival looks durable at current trajectory. Tier 3: Application layer products Every AI writing tool, image editor, productivity assistant, or vertical SaaS built on top of foundation models. These products live or die based on their distribution and product-market fit, not their infrastructure. The risk is that foundation models become commodities and eat upward into the application layer, as ChatGPT already has with basic writing, search, and coding. |
* * *
What this means for you
Why solo operators should be watching the compute war closely
If you are a solo operator or small team building on AI tools right now, the compute war affects you in at least three concrete ways. First, it is the underlying reason AI costs have been falling. The race to win infrastructure market share creates competitive pricing pressure that flows downstream to the tools you use. That trend continues as long as the infrastructure race does.
Second, it explains which AI tools are likely to disappear. When a product cannot secure favorable compute relationships, it faces margin pressure that either forces price increases or kills the product entirely. If a tool you depend on has no visible infrastructure backing, that is a risk worth factoring into your stack decisions.
|
Three practical questions for evaluating your AI stack Who is the underlying compute provider? Tools built on AWS, Azure, or Google Cloud with multi-year agreements are structurally more stable than tools built on spot capacity or unclear infrastructure relationships. Is the product sticky enough to survive commoditization? If the core feature of a tool is something that GPT or Gemini will offer natively within two years, the tool has a countdown on it. Are falling prices an opportunity? As infrastructure wars keep driving down model costs, the economics of building agent pipelines, automating repetitive tasks, and processing large volumes of data keep improving. The time to invest in building those systems is while the cost curve is still falling. |
* * *
The real stakes of the infrastructure bet
There is a version of this story where the compute war is just spectacle: enormous numbers, enormous egos, and enormous data centers that primarily benefit shareholders and analysts. That version is not wrong, but it is incomplete.
The infrastructure decisions being made in 2026 will determine what is technically possible in 2028. How capable the models are. How cheap it is to run an AI agent continuously. Whether autonomous workflows become accessible to individuals or remain the exclusive domain of large enterprises with large budgets. The companies spending hundreds of billions are betting that AI becomes as foundational as electricity. If they are right, whoever owns the power grid owns the future.
For solo operators and small teams, the most useful frame is not to try to predict who wins the infrastructure war. It is to build in a way that lets you move across providers, stay on top of the model and pricing changes that infrastructure competition creates, and treat falling compute costs as the tailwind they actually are.
|
The infrastructure war is not being fought over who has the best AI. It is being fought over who controls the cost of intelligence itself. That is a different and much more consequential competition. |
* * *
|
Before you go Do you think the compute war ends in a monopoly, a two-player duopoly, or does something disrupt it from below? Hit reply and share your read on how this plays out. The sharpest responses will shape a follow-up issue on what a world where compute is cheap and abundant actually looks like for everyday AI users. |
Until next time,
AI Spotlight
Practical AI, translated into real work, once a week.