The AI Spending Gap Is No Longer a Forecast

Meta's Q2 numbers landed last week with a figure that captures the whole AI spending problem in a single line. Profits dropped 14% to $6 billion. The return on the infrastructure spending that erased those profits was a 15-basis-point increase in Instagram sessions ^[1].

This drove a 15-basis point increase in sessions on Instagram, with particular strength in reshares and time spent, which are both strong indicators of better content-to-user matching. — Pavel Durov

That ratio — billions out, basis points back — is no longer a risk analysts warn about. It is the operational condition the industry is now navigating, and the evidence that it has become operational is not in any one earnings miss but in the fact that the largest AI companies are all pivoting in the same three directions at once. The first pivot is the most direct: from frontier performance to cost-discovery. OpenAI slashed GPT-5.6 Luna API pricing by 80% to 20 cents per million input tokens in late July, explicitly citing a shift from maximizing usage to what it called "valuemaxxing" — a term that would have been unthinkable from the company six months ago ^[2].

I think we’ll have a lot of ways we can help people get more value for less spend. — Sam Altman

The price cut was a direct response to Chinese open-weight competition from Kimi K3, DeepSeek V4, and GLM-5.2, but the competitive pressure only explains the timing. The deeper shift is in what OpenAI is optimizing for: not benchmark scores but the price at which enterprise customers will actually buy. SAP's CFO Dominik Asam made the same point from the buyer's side, stating plainly that the most advanced models are not always the best choice — enterprises will prioritize cost-effective, reliable tools over frontier performance ^[3].

It requires much more excruciating assurance levels. — Dominik Asam

The enterprise numbers bear this out. Corporate AI spending is up 110%, but ServiceNow's maturity index puts readiness at 51 out of 100 — the data infrastructure cannot keep pace with the tooling ^[4]. Meanwhile, AI token costs are blowing through annual budgets in months: Uber exhausted its AI coding budget before the year was half over, Microsoft restricted engineers from using third-party coding tools, and Goldman Sachs forecasts a 24-fold increase in monthly token consumption by 2030 — a trajectory that makes the cost pressure persistent, not a one-time shock ^[5]. A Saviynt executive put it bluntly: long-term cost models for AI currently lack a logical basis. Then there is Apple, running what amounts to a controlled experiment on whether the $900 billion capex arms race is necessary at all ^[6]. The company spent $12.7 billion on capex — against a collective $700 billion-plus from Alphabet, Amazon, Microsoft, and Meta — using a hybrid model of rented cloud and in-house chips. It generated a record $28 billion in operating cash flow and 17% revenue growth ^[7]. Apple is not abstaining from AI. It is testing whether AI monetization requires AI infrastructure ownership, and the early returns suggest it may not. The second pivot runs through geography. In the span of a single week, two major AI infrastructure deals landed in India, both framed around data sovereignty rather than model capability. Anthropic launched in-country Claude inference via AWS Bedrock, targeting regulated sectors where keeping data within national borders is, in the company's own words, the prerequisite to move AI from pilots into production ^[8].

India's banks, insurers, telecoms, public and government agencies steward the data of a billion people. When that data can stay in India, AI moves from pilots into the systems that matter most. — Irina Ghose

Days later, IBM and Sarvam AI announced a partnership to build sovereign AI infrastructure for Indian government and regulated enterprises, with models trained within India and data kept under Indian regulatory control ^[9]. The pattern extends well beyond India. Japan and NVIDIA launched a national AI infrastructure project with 27,500 Rubin GPUs, with Jensen Huang framing it in explicitly national terms: Japan must own, improve, secure, and deploy Japan AI — the country cannot outsource its national intelligence ^[10].

Japan must own, improve, secure, and deploy Japan AI. — Jensen Huang

China's Zhipu AI launched a 1-gigawatt data center powered exclusively by Chinese-made silicon, dedicated to training next-generation models independent of NVIDIA hardware ^[11]. Each of these moves is different in its particulars — national pride, regulatory compliance, supply-chain security — but they converge on the same operational logic: the global cloud model that assumed data would flow to wherever the best models lived is giving way to a model in which infrastructure follows jurisdiction, because enterprise revenue cannot be unlocked any other way. The third pivot is the one the market is now enforcing: from build-it-and-they-will-come to prove-the-return. Alphabet raised capex guidance past $200 billion and reported its first-ever cash-flow negative quarter. A four-day selloff in late July erased $1.2 trillion from U.S. tech stocks, with investors expressing what one account described as skepticism regarding the timeline and certainty of profits from the massive AI investments ^[12][13].

A lot of the demand is deferred, not destroyed. — Arvind Krishna

Meta's 14% profit drop and the 15-basis-point Instagram gain is the same story in miniature. The spending is real. The earnings are not. And the market, which was willing to fund the buildout on the promise of future returns, is beginning to ask to see the returns now. The most striking signal may be the one that involves no public market at all. SoftBank has committed over $60 billion to AI, including its OpenAI stake, with $30 billion in obligations due in the second half of 2026. It funded the push with a $40 billion bridging loan and a $20 billion margin loan against its Arm holdings. But lenders have refused to accept SoftBank's OpenAI holdings as collateral ^[14]. A lender will lend against Arm shares — a traded security with a daily price. It will not lend against a private AI company's paper valuation, no matter how high. That is not a forecast. It is a credit committee's operational judgment about what AI equity is worth as security, rendered in real time. The counter-evidence is real and worth taking seriously. NVIDIA reported 85% revenue growth and authorized an $80 billion share buyback ^[15]. The company is selling the picks and shovels, and the gold rush — whatever its ultimate economics — is generating enormous revenue for the suppliers. JPMorgan and Morgan Stanley argue the selloff may be overdone, noting that hyperscaler earnings are changing minds on AI ROI ^[16]. Bank of America's Vivek Arya contends that closed models still outperform open-weight alternatives in complex reasoning and agentic tool-use, and that enterprises prioritize managed infrastructure over cost ^[17]. But all of this describes the supply side — the hardware demand, the infrastructure buildout, the model performance benchmarks. It does not answer the question the three pivots are all responding to: whether end-user applications generate enough revenue to justify $900 billion in annual spending ^[6]. NVIDIA's revenue is evidence that the hyperscalers are buying chips. It is not evidence that the hyperscalers will earn back what they spent on them. What is visible now is not a bubble popping but something more interesting: an industry running cost-discovery experiments on itself after the money is already committed. These companies are not behaving like builders of an inevitable future. They are behaving like operators who have placed an enormous bet and are now, in real time, trying to find out whether the economics work. SoftBank's lenders have already rendered a partial verdict — one that required no analyst report, no earnings call, no market selloff. Just a loan officer looking at the collateral and saying no.