Five weeks ago, we published The $650 Billion Squeeze. The argument was straightforward. Big Tech was writing checks that rivalled national GDPs to build AI infrastructure. The software companies sitting on top of that infrastructure were watching their valuations evaporate. And the market was splitting into two tiers: infrastructure winners and software losers.

Since then, every major data point has arrived. Nvidia reported earnings. Anthropic sued the Pentagon. Microsoft announced it would embed Claude into its office suite. Oracle emerged as a surprise infrastructure contender. Mastercard launched the first authenticated AI agent transactions. PayPal installed a new CEO. And the term "SaaSpocalypse" went from a blog post to a CNBC headline.

The squeeze did not ease. It accelerated.

The $660 billion figure we cited in February was already outdated by the time we published it. Multiple sources now place combined 2026 hyperscaler capex closer to $700 billion.

The individual commitments have not changed. Amazon at $200 billion. Alphabet at $175 billion to $185 billion. Meta at $115 billion to $135 billion. Microsoft at approximately $145 billion. Approximately 75 percent of that aggregate spending is AI-specific, representing roughly $450 billion flowing directly into AI infrastructure in a single year.

What has changed is the arrival of a fifth player.

Oracle reported Q3 FY2026 earnings on March 10. Cloud infrastructure revenue jumped 84 percent to $4.9 billion. AI infrastructure revenue surged 243 percent year over year. Remaining performance obligations quadrupled to $553 billion. The stock jumped 10 to 15 percent. Oracle announced plans to raise $45 billion to $50 billion in FY2026 for cloud infrastructure expansion, targeting 10 gigawatts of computing power over three years.

Oracle was not part of the original $650 billion conversation. It is now. Add its commitments and the combined hyperscaler infrastructure spend approaches $750 billion.

Apple remains the strategic outlier. Projected FY2026 capex of just $14 billion, essentially flat year over year. But Apple is sitting on more than $130 billion in cash, positioned for acquisitions if AI startup valuations decline. Its 250,000 square foot server manufacturing facility opens in 2026. Apple is not ignoring AI. It is waiting for the right moment to deploy capital on its own terms.

In February, we flagged Nvidia's earnings as "the single most important data point for the entire AI infrastructure thesis." On February 26, the data arrived.

Q4 FY2026 revenue: $68.1 billion. Up 73 percent year over year. Up 20 percent quarter over quarter. A record. Full fiscal year revenue: $215.9 billion, up 65 percent. Earnings per share: $1.62, up 82 percent. Gross margin: 75 percent, beating guidance.

Then the Q1 FY2027 outlook: $78 billion in expected revenue, plus or minus two percent. Wall Street consensus was $72 billion. Nvidia beat it by $6 billion before the quarter even started.

Data centre revenue alone climbed 75 percent. The hyperscalers are not just announcing spending. They are spending. And the receipts flow through Nvidia.

This result matters beyond the stock price. In February, the bear case for AI infrastructure was that $660 billion in capex commitments might represent irrational exuberance, spending driven by competitive anxiety rather than actual demand. Nvidia's numbers say otherwise. The demand is real, it is growing, and the infrastructure layer is converting every dollar of hyperscaler commitment into revenue at historically high margins.

The infrastructure bet is not a bubble. It is a buildout. And the buildout is accelerating.

When we published in February, the software selloff was a week old. Thomson Reuters had fallen 15 percent. LegalZoom had dropped nearly 20 percent. Short positions in software stocks reached $24 billion.

Five weeks later, the selloff has a name. TechCrunch called it the SaaSpocalypse. The data justifies the label.

Forward price-to-earnings ratios across the SaaS sector have fallen from 39x to 21x. Software sector earnings expectations turned negative for the first time since 2022. The iShares Expanded Tech-Software Sector ETF is down more than 20 percent in 2026.

The individual stories are stark. HubSpot at $269, down 39 percent year to date, following a 42 percent decline in 2025. Figma down approximately 40 percent. Atlassian down 35 percent. Shopify down 29 percent. Salesforce at $195, down from a $296 peak. LegalZoom near its 52-week low of $6.45.

But the narrative is no longer one-directional. Something interesting happened in late February.

Thomson Reuters announced that its AI legal assistant, CoCounsel, had reached one million users. The stock surged 16 percent in a single week. Revenue guidance was raised to 7.5 to 8 percent growth. Salesforce reported that its AI agent platform, Agentforce, had hit $800 million in annual recurring revenue, up 169 percent year over year, with 29,000 deals closed.

Wedbush called the selloff "overblown." The partial recovery followed Anthropic's announcement of enterprise partnerships, which triggered a bounce across multiple software names.

Here is our updated read. The SaaSpocalypse is real, but it is not uniform. The software companies that are embedding AI agents into their own products are recovering. The ones that are not are continuing to fall. The market is no longer asking "will AI disrupt SaaS?" It is asking "which SaaS companies will survive the transition?"

The per-seat pricing model is the casualty. When AI agents compress the number of humans needed to do knowledge work, the number of seats a company buys compresses with it. The companies that can shift to outcome-based or usage-based pricing will adapt. The rest will be repriced indefinitely.

In February, we wrote about Anthropic's Cowork desktop agent triggering the worst software stock selloff in two years. That was the opening act.

On March 9, Microsoft announced Copilot Cowork, integrating Anthropic's Claude directly into Microsoft 365. A new E7 licensing tier, priced at $99 per user per month, launches May 1. It includes Copilot Cowork at $30 per user per month, giving enterprise customers Claude alongside OpenAI's models inside the applications they already use.

This is the most significant development since our original article, and it was not on anyone's radar five weeks ago.

Microsoft is now offering Claude and GPT inside the same productivity suite. The company that invested billions in OpenAI is hedging that bet by partnering with OpenAI's biggest competitor. For Microsoft, the AI model is becoming interchangeable. The platform is what matters.

For the SaaS sector, the implications are severe. If the world's dominant enterprise software company is bundling AI agents at $30 per seat on top of its existing Office suite, what is the value proposition of standalone AI-powered legal tools, marketing platforms, or financial analysis software? The standalone SaaS application increasingly looks like a feature inside Microsoft 365.

Meanwhile, Anthropic's trajectory has been extraordinary. Revenue has grown from $1 billion in December 2024 to approaching $20 billion in early March 2026. The company closed a $30 billion funding round at a $380 billion valuation, the second-largest private tech round ever.

And then Anthropic sued the Pentagon.

The Department of Defence classified Anthropic as a supply chain risk to national security, using authority designed for foreign adversaries, because Anthropic refused to strip safety restrictions from Claude for military use. The specific demands: mass domestic surveillance capabilities and integration into autonomous lethal weapons systems.

The response was unlike anything the AI industry has produced. Microsoft filed an amicus brief. 37 employees from OpenAI, Google, and DeepMind, including Google Chief Scientist Jeff Dean, signed a support letter. 22 former senior military and intelligence officials backed Anthropic. The EFF, FIRE, and Cato Institute filed First Amendment briefs.

This is a company approaching $20 billion in revenue, backed by a $380 billion valuation, drawing support from its direct competitors, while simultaneously fighting the federal government over the right to build AI safely. The outcome will shape AI governance for years.

OpenAI is no longer just building models. It is building the enterprise platform that sits on top of them.

The company raised $110 billion at a $730 billion valuation in February, with Nvidia and SoftBank each contributing $30 billion and AWS providing $50 billion. Annualised revenue stands at $25 billion, up from $20 billion at the end of 2025. The latest model, GPT-5.4, launched in March.

But the strategic story is Frontier, OpenAI's enterprise platform for building, deploying, and managing AI agents. Confirmed customers include HP, Intuit, Oracle, State Farm, Thermo Fisher, and Uber. Pilots are running at BBVA, Cisco, and T-Mobile.

The Verge described Frontier as "HR for AI," giving agents shared context, onboarding, permissions, and governance. That framing is useful. OpenAI is not just selling intelligence. It is selling the management infrastructure that makes intelligence deployable at enterprise scale.

The parallel with Microsoft's Copilot Cowork is instructive. Both companies are converging on the same insight: the value is not in the model alone. It is in the orchestration layer that connects models to enterprise workflows, data, and governance. The platform companies are consolidating fast.

In February, we noted that hyperscale data centres were straining existing grids. Five weeks later, the numbers have sharpened into something closer to an emergency.

Global data centre electricity consumption is projected to exceed 1,000 terawatt-hours by 2026, doubling from 460 terawatt-hours in 2022. Data centre power demand is growing at approximately 15 percent annually, four times faster than all other sectors. Occupancy rates are projected to hit 95 percent in late 2026.

Goldman Sachs estimates $720 billion in grid spending is needed through 2030. Researchers at Carnegie Mellon project that data centres could raise average US electricity bills by 8 percent by 2030, exceeding 25 percent in northern Virginia, the densest data centre market in the world.

The nuclear arms race we flagged in February has escalated. Microsoft's 20-year, $16 billion power purchase agreement with Constellation Energy for 835 megawatts from the renamed Christopher M. Crane Clean Energy Center (formerly Three Mile Island Unit 1) is expected online in 2028. Amazon has secured a 1.92 gigawatt agreement from Susquehanna plus $500 million in small modular reactor development. Google has contracted a fleet of SMRs through Kairos Power, targeting 2030.

The new entrant is Meta. The company announced "Prometheus," an AI data centre project requiring 6.6 gigawatts of nuclear procurement. For context, 6.6 gigawatts is roughly the output of six large nuclear power stations.

The AI infrastructure buildout is no longer just a technology story. It is an energy policy story, a real estate story, and increasingly a political story. The AI Infrastructure Act of 2025 provides tax credits for domestic data centre construction and streamlined permitting for small modular reactors. The regulatory environment is being reshaped around the assumption that this spending is not slowing down.

There is a story that did not exist five weeks ago, and it complicates the infrastructure thesis in ways that deserve attention.

Amazon is spending $200 billion on AI infrastructure in 2026, the largest capital commitment in company history. It is also discovering, in real time, that AI-generated code can take down its own services.

Amazon's AI coding tool, Kiro, caused at least two significant outages. A 13-hour AWS outage in December 2025 when Kiro deleted and recreated a production environment. A six-hour Amazon.com shopping outage on March 5 when AI-assisted code changes cascaded through the retail platform.

SVP Dave Treadwell convened an urgent engineering meeting on March 10, telling staff that site availability "has not been good recently." The new policy: junior and mid-level engineers must get senior sign-off on all AI-assisted deployments.

Here is the paradox. Amazon cut approximately 30,000 corporate roles since October 2025, with 40 percent targeting engineering. It set an 80 percent AI tool adoption target tracked as a corporate OKR. And now it is adding human checkpoints back into the process because the AI tools are causing production failures that the reduced engineering team cannot catch fast enough.

This is not an argument against the infrastructure buildout. It is a data point about the gap between deploying AI infrastructure and deploying AI safely. The company spending the most on AI is also the first to discover that the deployment pipeline was not designed for the speed and volume at which AI produces changes.

Gartner projects that 40 percent of agentic AI projects will be cancelled by the end of 2027. Forrester predicts at least two major multi-day hyperscaler outages in 2026. Amazon may have already delivered one of them.

Five weeks of data have sharpened the sorting we described in February. The market is now organised into five categories.

Infrastructure winners are having their best run in history. Nvidia's $68.1 billion quarter validated the entire thesis. Oracle emerged as a surprise contender with 243 percent AI infrastructure revenue growth. TSMC continues to benefit as the sole manufacturer of cutting-edge AI chips. Cerebras, backed by Benchmark Capital's $225 million special fund, is the challenger to watch.

Platform winners are being valued like countries. Anthropic at $380 billion. OpenAI at $730 billion. Alphabet above $4 trillion. Microsoft embedding both Claude and GPT into its enterprise suite. The platform layer has consolidated faster than anyone predicted.

SaaS survivors are the new category that did not exist in our original analysis. Thomson Reuters recovered 16 percent after CoCounsel hit one million users. Salesforce's Agentforce reached $800 million ARR. These companies are not being destroyed by AI. They are absorbing it. The survivors are the ones that moved fast enough to become AI-native before the repricing was complete.

SaaS casualties continue to fall. LegalZoom near its 52-week low. HubSpot down 39 percent year to date. The WisdomTree Cloud Computing Fund reflects the divide. DigitalOcean, an infrastructure play, is up 12 percent in 2026. The rest of the software index is down 24 percent. Infrastructure wins. Software-as-a-service, increasingly, does not.

Energy and physical infrastructure remains the most underappreciated category. Power companies, cooling technology providers, and data centre real estate firms are benefiting from a capex cycle with no historical precedent. Goldman Sachs' $720 billion grid spending estimate through 2030 represents an entirely new investment category that barely existed 18 months ago.

For our payments-focused readers, the five weeks since our original article have produced three developments that deserve close attention.

First, Mastercard has moved from announcement to execution faster than expected. Agent Pay is now live in Australia, representing the first authenticated agentic transactions processed through a major card network. US issuers are enabled. Global rollout is expected by the end of Q1. The Mastercard Agent Suite launches in Q2 to help merchants integrate agentic AI, backed by 4,000 advisors.

Most significantly, Mastercard unveiled Verifiable Intent on March 5, an open-source framework that creates cryptographic proof of what a user authorised when an AI agent transacts on their behalf. This aligns with Google's Agent Payments Protocol and Universal Commerce Protocol. The infrastructure for agentic commerce is being standardised in real time.

Second, PayPal installed Enrique Lores as CEO on March 1. The stock sits near $47, close to multi-year lows. Early signals include a $6 billion share buyback programme, a partnership with Sabre and Mindtrip for AI-powered travel booking, and a merchant-focused investment strategy. Lores grew HP's stock from $10 to $37 over six years. He has a similar distance to cover at PayPal, in a harder market.

Third, the banking sector is moving. 44 percent of finance teams will use agentic AI in 2026, a 600 percent increase. One financial services firm already has 60 agentic agents in production with plans for 200 more. Top use cases: fraud detection at 64 percent, loan processing at 61 percent, customer onboarding at 59 percent. AI agents have reduced credit turnaround times by 30 percent and purchase order cycle times by up to 80 percent.

But the gap between ambition and execution is enormous. 99 percent of companies plan autonomous agents in production. Only 11 percent have actually deployed them. The global AI agents market stands at $10.91 billion in 2026, growing at 49.6 percent annually to a projected $183 billion by 2033.

In February, we made four predictions. Here is how they landed.

Nvidia earnings would validate the infrastructure thesis. Correct. $68.1 billion in quarterly revenue, $78 billion in Q1 guidance, 75 percent margins. The infrastructure bet is not a bubble.

The capex-to-revenue gap would face scrutiny within 12 to 18 months. Too early to grade, but Amazon's free cash flow going negative is drawing exactly the analyst attention we predicted. The clock is ticking.

SaaS companies would be forced to choose: embed AI, become the compliance layer, or get compressed. Correct, and the sorting happened faster than we expected. Thomson Reuters and Salesforce chose to embed. Others are still deciding. The market is not waiting.

The payments and fintech sector would face the same repricing. Partially correct. PayPal's leadership reset and Intuit's stock decline are the early signals. But Mastercard's proactive move into agentic commerce suggests the card networks may be better positioned than the software layer to navigate this transition.

What we did not predict: Microsoft partnering with Anthropic to offer Claude inside its office suite, Oracle emerging as a fifth major infrastructure player, or the Pentagon classifying an AI safety company as a national security risk. The speed of change has exceeded even our aggressive projections.

Four things to watch in the weeks ahead.

First, the Microsoft E7 rollout on May 1. At $99 per user per month, this is the first real price point for enterprise AI agents bundled into existing productivity software. If adoption is strong, the repricing of standalone SaaS tools accelerates. If enterprises baulk at the cost, the SaaS survivors get more time.

Second, Amazon's deployment crisis. The company spending $200 billion on AI infrastructure is simultaneously adding human checkpoints to slow AI-assisted code deployment. How Amazon resolves the tension between its 80 percent AI adoption target and its reliability problems will be a template for every enterprise running AI at scale.

Third, Mastercard's global Agent Pay rollout. Authenticated AI agent transactions are live in Australia. The US is next. If agentic commerce reaches mainstream adoption through the card networks rather than through fintech startups, the payments landscape shifts permanently.

Fourth, the Anthropic v. Pentagon outcome. If Anthropic wins, it creates a legal shield for safety-oriented AI development. If the government wins, it sets a comply-or-lose-contracts precedent that chills safety research across the industry. This is the most consequential AI governance case in history, and it is happening now.

In February, Deutsche Bank 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." Five weeks later, the winners are winning bigger, the losers are losing faster, and the squeeze is tighter than anyone expected.