The Two-Way Squeeze Reshaping the Software Industry

"Learning on our dime." That was OpenAI CFO Sarah Friar explaining why the company is spending $115 billion to build its own data centers by 2029 ^[1]. Cloud providers had been the middlemen, and the model lab had decided to cut them out.

What I want to make sure is that we're not giving that IP away. — Sarah Friar

The candor was unusual, but the direction of travel was familiar. In the same months that OpenAI laid plans to move up into physical infrastructure, it and its chief rival Anthropic were also moving down into applications. In April, Anthropic launched Claude Enterprise plugins that integrated directly into Excel, PowerPoint, and Slack ^[2]. The announcement triggered a sell-off across software stocks; 24 articles tracked the fallout ^[3]. Then in July, both labs launched standalone implementation companies designed to embed engineers directly inside mid-sized enterprises: OpenAI Deployment Company, a $4 billion joint venture with 150 forward-deployed engineers, and Ode with Anthropic, capitalized at $1.5 billion ^[4]. The model labs are integrating in both directions at once, and the traditional software industry is caught between them. The upward squeeze is the more visible half. OpenAI's $115 billion data center plan is one line item in a broader revaluation of everything physical that AI touches. Amazon raised its 2026 AI capital spending forecast to $220 billion, with CEO Andy Jassy saying the company "will still not have enough capacity to meet all the demand we have in 2026" ^[5]. Meta forecast $125 to $145 billion in AI data center spending for the year and launched Meta Compute to lease excess capacity to external customers ^[6]. Nvidia's Jensen Huang projects the AI infrastructure market will reach $3 to $4 trillion by 2030, with the top four hyperscalers already spending roughly $600 billion annually ^[7]. The physical premium has spilled beyond chips and server racks. Energy Transfer secured a 20-year gas supply agreement to power Meta's Louisiana data center and a deal with Oracle for 900,000 Mcf per day across three facilities, contracting over 6 billion cubic feet per day of new pipeline capacity ^[8]. Blackstone raised $1.75 billion in an IPO for its Digital Infrastructure Trust, a data center REIT with $25 billion in identified near-term acquisition targets ^[9]. Switch Inc., a Las Vegas data center firm, confidentially filed for an IPO that could value it at nearly $50 billion ^[10]. Memory-chip makers SK Hynix and Micron both crossed $1 trillion in market capitalization in May, driven by high-bandwidth memory demand expected to exceed supply for three years ^[11]. The downward squeeze is quieter but just as consequential. When UBS analyst Karl Keirstead downgraded ServiceNow from Buy to Neutral in April, cutting his price target from $170 to $100, his rationale was specific: Fortune 500 enterprises are constraining spending on non-AI software to prioritize "data and infrastructure investments expected to inflect in 2026" ^[12].

Beginning in December, we began hearing [Fortune 500] enterprises and partners express a view that because AI and the associated data and infrastructure spend were expected to inflect in 2026, spending on non-AI or core software spend was now under greater pressure. — Karl Keirstead

The numbers behind the downgrade were stark. In a single quarter, Salesforce dropped 26%, Adobe fell 29.7%, and Snowflake declined 21% ^[13]. ServiceNow's stock fell nearly 50% from its record high ^[14]. Airtable's valuation collapsed from $12 billion to $1.3 billion ^[15]. The market was drawing a line between companies that own physical infrastructure and companies that rent it. The model labs' downward move sharpens this line. Anthropic's Claude now writes 80% of its own code, yielding an 8x increase in code production per person ^[16]. Claude Enterprise plugins do not just assist workers; they replace the workflows that standalone SaaS companies used to sell. When the model itself can handle the investment banking analysis or the HR onboarding flow, the software that used to charge a subscription for that workflow becomes redundant. The labs' new implementation ventures take this further: instead of selling a model through an API and letting a third-party consultancy or SaaS firm handle deployment, OpenAI and Anthropic are embedding their own engineers directly inside enterprises ^[4]. The implementation layer is being reframed from an independent value creator into a distribution channel for the physical stack. SaaS companies are not passive in this. They are pursuing three distinct survival strategies, and all three require subordinating to the physical layer in some form. The first is becoming AI-native by integrating frontier models directly. Snowflake signed a $200 million deal with Anthropic to integrate Claude into its Cortex AI platform, positioning itself as the governed-data layer for AI ^[17]. Salesforce's Agentforce product saw ARR surge 330% to $540 million, and the company raised its 2027 revenue guidance to $45.9 to $46.2 billion ^[18]. Palantir grew revenue 63%, with US commercial revenue up 121% ^[19]. These firms are not fighting the models. They are wiring themselves into them. The second strategy is becoming the governance layer. ServiceNow's CEO Bill McDermott has positioned the company as model-agnostic, a platform that manages AI agents regardless of which lab built them. "We integrate with any model, cloud, interface, data, and system they choose to deploy," he said ^[20]. The company reported 130% year-over-year growth in $1 million-plus Now Assist customers and a $28 billion backlog, even as the stock traded 42% below its peak ^[20]. McDermott bought $3 million in company stock to signal his conviction ^[14]. The bet is that enterprises will need a neutral layer to govern the sprawl of AI agents, and that this layer sits above the models but below the applications. The third strategy is the most radical: become infrastructure yourself. Meta Compute is the clearest example, a platform company pivoting toward selling compute access to external customers ^[6]. Akamai Technologies, historically a content delivery network, is now investing roughly 40% of revenue in capex for distributed cloud infrastructure targeting edge AI and robotics, suspending share buybacks to fund the buildout ^[21]. SpaceX's record $85.7 billion IPO included an explicit pivot to AI infrastructure, acquiring xAI and securing compute contracts with Anthropic, Alphabet, and Reflection AI to rent access to its Colossus data center ^[22]. These are companies that could have remained in the application or services layer and chose instead to compete for a piece of the physical stack. The pattern is consistent enough to name: the implementation layer is being squeezed from above by models that can do the work and from below by infrastructure costs that consume the budget. But the pattern is not a settled equilibrium. The counter-evidence is substantial. On the infrastructure side, the buildout itself is stalling. Of 3,969 data center facilities announced, only 802 are under construction, constrained by shortages of specialized labor, building materials, and TSMC chips ^[23]. New York and Texas have imposed building moratoriums. The Federal Energy Regulatory Commission ordered that residential ratepayers not subsidize data center power costs. Michael Burry has shorted Oracle and Nebius Group, arguing that companies are understating GPU depreciation by extending the useful life of assets.

one of the more common frauds of the modern era. — Michael Burry

On the application side, the "SaaSpocalypse" narrative is incomplete. Atlassian revenue grew 32% to $1.79 billion, with its AI Service Collection exceeding $1 billion in ARR ^[24]. Twilio and Five9 also posted strong earnings in the same quarter that punished Salesforce and Adobe ^[24]. The bifurcation may be within SaaS — between firms that successfully reposition as AI-native and those that do not — rather than a uniform discount applied to the entire implementation layer. RBC Capital Markets has framed the moment explicitly as a replay of the cloud transition: software companies will see "growth and margins meaningfully fall (in many cases going negative) before reaccelerating, thus creating stronger companies" ^[25].

Similar to how software companies went through cloud transitions (some more belatedly than others), whereby we saw growth and margins meaningfully fall (in many cases going negative) before reaccelerating, thus creating stronger companies, we expect the same with AI — RBC Capital Markets

That is the honest read. The squeeze is real, and the mechanism is not market sentiment or a cyclical rotation. It is the stated strategic intent of the two most important model labs, executed in both directions at once. OpenAI and Anthropic are building the data centers that host the intelligence and the implementation teams that deploy it. What sits between them — the software industry that spent two decades building independent value on top of someone else's infrastructure — is being told, for the first time, that the infrastructure owner is also the application provider. The transition will be brutal. Whether it is permanent depends on whether the physical buildout can actually deliver the capacity the model labs have promised, and whether the SaaS firms now wiring themselves into those models can find margins on the other side.