Hyperscaler capex has hit levels that rival national economies. The software companies those dollars are meant to replace are paying the price.

In the span of 72 hours this week, Amazon committed to spending $200 billion on AI infrastructure, and Anthropic's Cowork desktop plugins triggered the worst software stock selloff in two years. One sector's investment thesis became another's existential crisis.

The numbers from earnings week paint a picture of an industry splitting in two. The companies building AI infrastructure are writing checks that rival national GDPs. The software companies sitting on top of that infrastructure are watching their valuations evaporate in real time. And in between, an entire ecosystem of energy companies, chip manufacturers, and platform players is being reorganised around a single question: who captures the value that AI creates, and who gets displaced by it?

Ben Thompson at Stratechery called it "SaaSmageddon." The data suggests the label fits.

Start with the capex numbers. Four companies, in the span of a single earnings week, committed to a combined approximately $660 billion in 2026 AI infrastructure spending.

Alphabet led with $175 billion to $185 billion, up from roughly $75 billion in 2025, a jump of more than 130 percent. Amazon followed with $200 billion, its largest capital commitment in company history. The figure is so large that analysts project Amazon's free cash flow will swing from surplus to a deficit of between $17 billion and $28 billion. This is a company that generated $36 billion in free cash flow in 2024, now voluntarily burning cash to build AI capacity.

Meta pledged $115 billion to $135 billion. Microsoft committed approximately $145 billion.

The combined total represents a jump of more than 60 percent from the $381 billion these four companies spent in 2025. For context, $660 billion exceeds the annual GDP of Sweden. It is roughly equal to the entire defence budget of the United States. These are not incremental increases. They are bets of a magnitude that the technology industry has never made before.

Then there is Apple, which committed just $12 billion in capex. But Apple's restraint is strategic, not accidental. After outsourcing its AI infrastructure to Google through the Gemini/Siri partnership, Apple posted record quarterly revenue of $144 billion while spending a fraction of what its peers burn on data centres. Apple is not ignoring AI. It is letting others build the infrastructure and collecting rent on top of it. The AI dividend, it turns out, can flow to companies that choose partnerships over picks and shovels.

Despite strong earnings across the board, the three biggest spenders lost a combined $900 billion in market value in the days following their announcements. Investors are asking a straightforward question: when does $660 billion in spending start generating $660 billion in returns?

The capex story might have dominated the week if not for what happened on the software side.

Anthropic released industry-specific plugins for its Cowork desktop agent, covering legal, finance, and marketing workflows. On the same day, it launched Claude Opus 4.6 with a one-million-token context window and multi-agent team capabilities. The market's reaction was immediate and brutal.

Thomson Reuters fell more than 15 percent. LegalZoom dropped nearly 20 percent. Salesforce is now down 25 percent year to date. The WisdomTree Cloud Computing Fund has lost 20 percent of its value in 2026. The S&P Software and Services Index dropped for eight consecutive sessions, shedding roughly 20 percent from its October peak.

Hedge funds moved fast. According to SaaStr, short positions in software stocks reached $24 billion. Microsoft alone lost $360 billion in market cap in a single day. Public B2B software stocks are down 30 to 40 percent in five weeks. The selloff was not limited to niche players. This was a broad repricing of what the market believes knowledge-work software is worth in an AI-native world.

The fear is not abstract. When a desktop AI agent can draft legal briefs, generate financial models, and build marketing campaigns from a single interface, the value proposition of purpose-built SaaS tools starts to look different. Not worthless, but worth less.

And the capability is not theoretical. Relevance AI, profiled this week by SaaStr, reported that 40,000 new AI agents were created on its platform in January alone. Canva, Databricks, KPMG, and Autodesk are already running production AI agents on it. Ops teams, not engineers, are building these agents. The abstraction layer that separated "users" from "builders" is dissolving.

Meanwhile, Sam Altman added fuel to the fire, predicting that AI agents will integrate any service they want, with or without official APIs. If that prediction holds, the moat for many SaaS companies is not just narrowing. It is being routed around entirely.

Not everyone agrees on what this repricing means. The analyst community is as divided as we have seen it in years.

The bear case: structural disruption. Constellation Research has warned of outright SaaS cannibalization. Rolf Bulk at Futurum argues that AI will "pressure profits and limit pricing power" across the software sector, particularly for companies whose core value is organising and displaying information that an AI agent can now synthesise on the fly. Platformer framed the question bluntly: will AI kill SaaS? The publication went further, testing whether a state-of-the-art chatbot could replace an entry-level journalism role entirely. The answer was nuanced, but the experiment itself is a signal of where the conversation has moved.

The ICONIQ State of AI Report, surveying approximately 300 executives at AI-focused software companies, found that while gross margins are improving, pricing models remain unsettled. R&D spending is surging. Revenue is growing. But nobody has figured out sustainable unit economics for AI-native products yet. The AI software layer is expanding, but it is expanding into territory that SaaS currently occupies.

The bull case: overreaction. Mark Murphy at JP Morgan called the market's extrapolation "illogical," arguing that enterprise software procurement does not move at the speed of a stock selloff. Dan Ives at Wedbush made a similar point: enterprises cannot "snap their fingers" and migrate decades of workflow tooling to an AI agent, no matter how capable it is. Compliance requirements, data governance, and integration complexity create switching costs that protect incumbents, at least for now.

Gartner weighed in with a measured take, noting that Cowork is "not a replacement for SaaS applications managing critical business operations." The distinction matters. An AI agent that drafts a contract is different from the contract management system that tracks versions, enforces approval workflows, and maintains audit trails. The scaffolding of enterprise software is not just the document. It is the process.

Our read: the truth sits between these poles. AI agents expose how much knowledge work remains manual, creating real pricing pressure on tools that automate the simple parts. But they do not eliminate the need for structured data, compliance frameworks, and enterprise integrations overnight. The pressure is real. The extinction is not. At least not yet.

There is a story beneath the capex numbers that deserves attention: where does all this power come from?

MIT Technology Review devoted significant coverage this week to the growing intersection of AI infrastructure and nuclear energy. Hyperscale data centres are consuming electricity at rates that are straining existing grids, and the industry's answer increasingly involves next-generation nuclear plants.

The logic is straightforward. AI training runs and inference workloads require consistent, high-capacity power that solar and wind cannot reliably deliver at the scale needed. Microsoft has signed agreements to restart a unit at Three Mile Island. Google and Amazon have both invested in small modular reactor companies. The hyperscalers are not just building data centres. They are building the power plants to run them.

This creates a new investment category that sits adjacent to the AI infrastructure thesis. Power companies like Vistra and nuclear energy plays have surged. The AI capex wave is not just reshaping technology. It is reshaping energy markets, real estate, and even national grid planning.

The market is sorting itself into four clear categories, and the separation between them is accelerating.

Infrastructure winners are having a historic run. Nvidia captures roughly 90 percent of AI accelerator spending, and its earnings report on February 25 will serve as the next major catalyst for the entire sector. Benchmark Capital just raised $225 million in special funds to double down on Cerebras, the Nvidia challenger it has backed since 2016. TSMC continues to benefit as the sole manufacturer of cutting-edge AI chips. The picks-and-shovels trade is alive and well.

Platform winners are being valued like countries. Anthropic recently reached a $350 billion valuation. OpenAI is targeting $800 billion or more. Alphabet has crossed $4 trillion in market cap, buoyed by both its cloud AI business and the Gemini licensing revenue flowing from Apple. OpenAI's launch of Frontier, an enterprise platform for building, deploying, and managing AI agents, signals that the platform layer is consolidating fast. As The Verge noted, Frontier functions as something like "HR for AI," giving agents shared context, onboarding, permissions, and governance. The platform companies are not just building models. They are building the management layer that makes agents useful at scale.

SaaS losers, at least for now, span the horizontal software landscape. HubSpot has fallen 39 percent. Figma is down 40 percent. Atlassian has dropped 35 percent. Shopify has lost 29 percent. These companies are not failing. Their businesses are still growing. But the market is repricing their future growth potential in a world where AI agents can replicate significant portions of their functionality. The question is whether these companies can reinvent themselves as AI-native platforms, or whether they will be compressed into the infrastructure layer beneath the agent.

Energy and physical infrastructure represents the fourth category, and it may be the most underappreciated. Power companies, cooling technology providers, and real estate firms specialising in data centre campuses are all benefiting from a capex cycle that has no historical precedent. This is the "second derivative" of the AI trade: companies that do not build AI at all, but profit from the physical reality that AI requires electrons, space, and thermal management at enormous scale.

For our payments-focused readers, the implications of this week's developments are significant and accelerating.

PayPal fired CEO Alex Chriss after just two and a half years and appointed HP's Enrique Lores effective March 1, according to Finextra. The board cited execution "not in line with expectations," particularly in branded checkout. As Payments Dive reported, slow merchant adoption of PayPal's latest technology and lagging growth of legacy checkout services stymied growth plans. PayPal's share price is down 80 percent from five years ago. Q4 missed estimates, the stock dropped roughly 18 percent, and 2026 profit guidance pointed to a decline.

The timing is telling. PayPal's struggles are not just about execution. They reflect a broader compression in payments technology, where AI agents are beginning to intermediate the checkout experience itself. Payments Dive's analysis of how agentic commerce threatens traditional retail marketplaces connects directly to this thesis. If AI agents shop and pay autonomously, the branded checkout button becomes less relevant. The agent does not care about the payment interface. It cares about the outcome.

Mastercard is leaning into the shift rather than resisting it. Its announcement that it will deliver agentic AI capabilities to merchants by the end of June represents one of the first major moves by a card network to embed AI agents directly into the payment flow.

Meanwhile, banks face a dual authentication crisis as AI agents begin initiating autonomous transactions, approving payments, and freezing accounts in real time. Traditional security frameworks, built to verify human identities, cannot address the fundamental shift from verifying who is making a transaction to understanding what an agent intends to do. This is not a theoretical problem. Financial institutions are deploying these agents now, and the security architecture has not caught up.

Four things to watch in the weeks ahead.

First, Nvidia's February 25 earnings. This is the single most important data point for the entire AI infrastructure thesis. Nvidia captures roughly 90 percent of AI accelerator spending, and its results will reveal whether the $660 billion capex wave is translating into actual hardware revenue, or whether we are witnessing the early stages of an infrastructure bubble. Every infrastructure bull case depends on Nvidia continuing to ship at scale.

Second, the capex-to-revenue gap. The hyperscalers are spending at rates that require extraordinary returns. Amazon's free cash flow going negative is a particularly stark signal. If AI products do not generate proportional revenue within the next 12 to 18 months, the market will reprice the infrastructure bet the same way it repriced SaaS. The difference is that infrastructure repricing would be measured in trillions, not billions.

Third, the SaaS response. The most interesting question is not whether AI agents will displace some SaaS functionality. They already are. The question is how SaaS companies respond. Do they embed AI agents into their existing products, turning the threat into a feature? Do they pivot to become the compliance and governance layer that sits beneath the agent? Or do they get compressed into commodity infrastructure? The next two quarters of earnings calls will reveal which strategy each company is choosing.

Fourth, the payments and fintech ripple effects. Opus 4.6's financial analysis capabilities, combined with Mastercard's agentic commerce push and the authentication crisis in banking, suggest that the SaaS repricing is heading towards financial services next. PayPal's leadership reset is a warning sign. Intuit's stock decline is another. The companies that intermediate financial workflows are facing the same existential question that legal and marketing software faced this week: what happens when an AI agent can do most of what your product does, inside a single interface?

Deutsche Bank has framed the macro picture well: we have moved from a world where "every tech stock is a winner" to a "true winners and losers landscape." The $650 billion squeeze is the mechanism by which the market is sorting one from the other.

Related reading from Major Matters:

As AI infrastructure spending reaches levels that rival national economies, is the software industry facing a temporary correction or a permanent repricing of what knowledge work is actually worth?