The market calls it a “boom.” The balance sheets call it an “investment.” But if you look at the thermodynamics, what we are witnessing in 2026 is a hostage negotiation between software capital and physical reality. And reality is winning.
We are currently watching the Big Four tech giants—Microsoft, Amazon, Meta, and Google—commit to a collective capital expenditure of $650 billion in a single calendar year. This is a 60% increase year-over-year. To put that number in perspective, it is larger than the GDP of Sweden. It is a level of industrial mobilization that we typically associate with wartime economies, not peacetime software development.
But the most dangerous misconception circulating in Silicon Valley right now is that this spending is “growth.” It is not growth. It is a defensive moat made of silicon, copper, and megawatts. It is the price of admission to a game where the table stakes have risen from “knowing how to code” to “owning a nuclear power plant.”
The Capex Cannibalism
The narrative for the last decade of SaaS (Software as a Service) was simple: write code once, run it everywhere, enjoy near-infinite margins. That era ended the moment the first H100 cluster went online. The new era is defined by Capex Cannibalism—the necessity to devour your own free cash flow to feed a hardware beast that depreciates faster than a used car.
When Microsoft or Meta pours $100 billion into data centers, they aren’t just buying chips. They are buying energy futures. They are buying land. They are buying cooling water. The “cloud” has condensed into something heavy, hot, and incredibly expensive.
The terrifying part for the software ecosystem is the trickle-down effect of this capex. When the underlying infrastructure costs $650 billion a year to maintain and expand, the cost of compute cannot go to zero. It must go up. The “race to the bottom” for API pricing is a temporary illusion funded by venture capital subsidies. Eventually, the bill for the $650 billion must be paid. And it will be paid by the application layer.
If you are building a startup that relies on cheap inference to drive a low-margin SaaS tool, you are building a house on a landlord’s land while he is actively raising the rent. The “Intelligence API” is not a utility like electricity, regulated and stable. It is a scarce commodity, rationed by the highest bidder.
The Stargate Paradox: 10 Gigawatts of Hubris
Nothing exemplifies this new era better than the Stargate Project.
A $500 billion initiative backed by OpenAI, SoftBank, Oracle, and MGX. The goal? A single compute cluster consuming 5 gigawatts of power by 2028, scaling to 10 gigawatts shortly after.
Let’s pause on the physics of that. 10 gigawatts. That is roughly the power consumption of New York City. We are talking about building a machine that consumes as much energy as a major metropolis, solely for the purpose of matrix multiplication.
The comparisons to the Apollo program or the Manhattan Project are thrown around loosely, but they are understated. The Manhattan Project cost roughly $30 billion in today’s dollars. Stargate is an order of magnitude larger. It is the largest industrial project in human history dedicated to a single informational output.
The paradox of Stargate is this: The more we centralize compute into these gigawatt-scale behemoths, the less “democratized” AI becomes.
We were promised a future where AI would empower the individual. Instead, we are building a future where AI is generated by a few God-like machines, owned by a consortium of sovereign wealth funds and tech monopolies. If Stargate succeeds, it doesn’t just create AGI; it creates a centralized intelligence authority. The “Open” in OpenAI becomes a historical artifact.
This is not software engineering. This is civil engineering. It involves securing land rights, negotiating with nuclear reactor vendors, and likely, employing private security forces that rival small armies. The barrier to entry is no longer a GitHub repo; it is a seat at the table with energy ministers.
Sovereign Silicon vs. SaaS Slums
While the consortiums build Stargate, we see another trend emerging: Sovereign Compute.
Elon Musk’s xAI and its “Colossus III” cluster in Mississippi is the prototype. A 2 gigawatt single-site installation. This is not just a data center; it is a sovereign territory of compute. By vertically integrating power generation, cooling, and silicon, xAI is attempting to opt out of the public cloud rent-seeking model.
This bifurcation of the market is critical to understand. On one side, you have the Sovereign Lords—entities like xAI, Meta, and the Stargate consortium—who own the physical stack from the electron to the token. On the other side, you have the Sovereign Serfs—the thousands of SaaS startups and enterprise “AI wrappers” that rent intelligence from the Lords.
The Serfs are trapped in the “SaaS Slums.” They fight over the scraps of margin left after paying the API tax. They are vulnerable to every pricing change, every rate limit, every policy shift of the Lords.
The Sovereign Lords, meanwhile, are playing a game of Silicon Realpolitik. They know that in a world of Agentic AI, the entity that controls the inference capacity controls the economy.
The Agentic Inference Tax
This brings us to the final nail in the coffin of the old software model: Agentic AI.
The industry is pivoting from “Copilots” (which help you write an email) to “Agents” (which do the work for you). The difference sounds subtle, but infrastructurally, it is violent.
A Copilot interaction is a single turn: Prompt -> Response. Low compute, low latency requirement.
An Agentic workflow is a loop: Plan -> Reason -> Tool Use -> Observe -> Reflect -> Act. A single user request might trigger hundreds of inference steps.
This shift moves us from a “Retrieval” economy to an “Execution” economy. And execution is expensive.
If an agent needs to “think” for 5 minutes to solve a complex coding task or negotiate a supply chain contract, that is 5 minutes of H100 time. Who pays for that?
The “Agentic Inference Tax” means that the cost of doing business is directly tied to the cost of compute. As agents become more autonomous, they consume more compute. The $650 billion capex bill isn’t just for training; it’s to build the inference capacity for a world where software thinks before it acts.
This destroys the traditional SaaS margin structure. You cannot have 90% gross margins when your COGS (Cost of Goods Sold) scales linearly with the intelligence of your product. We are heading toward a world where software looks more like a services business—with 40% margins—because the “labor” is being done by silicon that must be paid for.
Strategic Implication: The Physicality of the Cloud
We must abandon the metaphor of the “Cloud.” It is too fluffy, too weightless.
What we are building is the Steel Sky. A rigid, heavy, industrial canopy of infrastructure that requires massive amounts of capital, energy, and water to sustain.
For the investor, the signal is clear: avoid the middle. The value accrues to the physical layer (Nvidia, TSMC, Energy Providers, Data Center REITs) and the application layer that owns a unique proprietary dataset (which is rare). Everything in between—the “infrastructure software,” the “MLOps tools,” the “generic model providers”—is going to be crushed by the capex weight of the giants.
For the engineer, the lesson is starker: code efficiency is about to matter again. In a world of infinite free compute, optimization was a luxury. In a world where compute is rationed by gigawatt-scale power plants, efficient code is a survival trait.
The Stargate is opening. It is magnificent, terrifying, and expensive. And it is not a gateway to a post-scarcity utopia. It is a monument to the fact that intelligence, like everything else in this universe, has a cost. And in 2026, that cost is $650 billion and climbing.
If you aren’t paying the bill, you are the product. If you aren’t building the generator, you are just waiting for the lights to go out.
The verdict: The software margin is dead. Long live the silicon margin.