Oracle Is Borrowing for AI and Paying for It With Payroll

Oracle has committed roughly $55.7 billion to AI spending and is taking on billions in debt to pay for it. The same company has eliminated 21,000 full-time jobs, spent most of a $2.1 billion restructuring budget, and is reportedly planning another round in September, with some teams facing double-digit percentage cuts.

Those look like two stories. We think they're one decision, seen from opposite ends of the balance sheet.

The layoffs of 2022 through 2024 came from demand: companies overhired during the pandemic and then corrected. What's happening in 2026 is about financing. Companies are moving money from operating expense to capital expenditure, turning salaries into servers, and cutting people is the fastest way to fund a data center buildout that investors are watching closely. If you work in tech, which side of that trade your job sits on matters more right now than any skills list.

Oracle is growing and cutting at the same time

Other companies are copying Oracle's sequence, so it helps to walk through it. A restructuring budget gets approved. Severance goes out as a one-time charge. The recurring payroll cost disappears from future quarters, and the cash that frees up helps service the debt taken on for infrastructure. The March 2026 cuts were pitched as efficiency. The September round, reported while the company carries a $55.7 billion AI commitment, is a lot harder to pitch that way.

What makes this different from a normal downturn is that Oracle isn't shrinking. Its AI infrastructure business is expanding fast, and that's exactly why it needs the cash. People are being cut from a growing company because growth in this cycle takes capital far more than labor. Revenue per employee goes up, mostly because the revenue is coming from leased compute. The remaining staff didn't suddenly get more productive.

Tech layoffs in 2026 are already close to the full-year 2025 total, according to tracking across TikTok, Microsoft, Meta, Oracle, Samsung, Zillow and others. Roughly 205,000 of those cuts have been tagged as AI-linked. We think that label is sloppy.

The financial mechanism behind it is not.

Where the money went in one week

You could trace where the redirected capital lands in a single week of headlines. It's going to fabs, servers and model companies, and very little of it is showing up in engineering org charts.

• TSMC approved a $29.44 billion capital budget to expand capacity for AI demand.
• Intel raised $20 billion in an upsized share sale, diluting shareholders to fund manufacturing rather than staffing.
• Foxconn posted a quarterly profit surge on AI server demand.
• Anthropic is in talks to buy Decart AI for roughly $6 billion, buying capability instead of building a team.
• Lovable, a vibe-coding startup, raised $400 million at a $13.3 billion valuation. That price assumes far fewer people will be writing production code in five years.

Gaming, marketing and satellite offices go first

The cuts have a pattern. Salesforce cut roughly 1,000 roles, concentrated in marketing, product and communications. Microsoft is set to cut nearly 5,000 jobs as Xbox downsizes. LinkedIn is shutting its Israeli R&D operation and laying off nearly everyone there. What these have in common is that they cost money without feeding the AI infrastructure story investors are pricing in.

In a normal year you could defend consumer gaming, a regional R&D outpost, brand marketing or internal comms. In a year when the company is borrowing to build compute, each of them is up against the data center line item, and they lose. LinkedIn's Israel closure is the one we'd study. Nothing we've seen suggests the site was underperforming. What counted against it was being a separate cost center that could be closed cleanly, inside a company whose parent is spending heavily somewhere else.

The jobs that survive sit close to the capital: infrastructure engineering, data platform work, model deployment, chip design, and the sales teams that fill new capacity. China's Kiwimoore, a chip designer, is heading to a Hong Kong IPO at a $2 billion valuation. That's the corner of the labor market still clearing at a premium.

Microsoft's president thinks the jobs apocalypse was overestimated

The executive justification is productivity: AI does more, so you need fewer people. Employees keep describing something else. Surveys this week found staff working up to 90 hours a week at companies whose leaders say AI means less work. Both can be true if the headcount gets cut on capex logic first and the productivity story gets written afterward.

Microsoft's president added a caveat that cuts against the industry's own pitch, saying AI leaders overestimated the jobs apocalypse. We think that's probably right, and it makes the layoffs harder to explain as automation. If AI isn't yet displacing labor at the task level in most functions, then capital allocation is what's driving the 2026 cuts. The work doesn't vanish. It gets spread over fewer people while the savings service infrastructure debt.

If you survive one of these rounds, don't mistake that for stability. You'll usually be absorbing the departed team's scope with the same tools you had last quarter.

Policy has started to notice the asymmetry underneath all this. Payroll is expensed. Equipment and infrastructure get depreciation and, in many jurisdictions, accelerated write-offs and local incentives on top. Representative Casar's proposed AI token tax is an early attempt to close that gap, pitched as ending what supporters call a payroll subsidy for automation, with the revenue going to a public jobs program. We have no idea whether that bill goes anywhere. Add French publishers pressing the antitrust regulator over Google's AI and Australia reworking its media law to make platforms pay outlets, and you have regulators starting to attach costs to AI deployment. That might slow the conversion down. We don't see it reversing the buildout.

If you work at a company building AI infrastructure, read the capital plan before you read the layoff memo. Quarterly capex guidance, new debt issuance and announced data center commitments tend to show up six to nine months ahead of the headcount decisions they cause. Oracle's September plans were visible in its spending disclosures long before any team got the calendar invite. In your own employer's earnings materials, a sharp jump in capex relative to revenue usually comes before a restructuring charge.

Then work out where your function sits relative to that spending. Roles that cut the cost of running compute or fill it with paying customers are on the funded side, and so is the work that makes AI legal to deploy. Pure overhead against the infrastructure story is exposed no matter how good your reviews are. Ask in a skip-level whether your team's budget sits in opex or belongs to a product with its own P&L. The skills we'd build are the ones near infrastructure economics: cost optimization, evaluation and reliability of AI systems, data engineering, and technical sales that deals with procurement.

Satellite offices carry extra risk, because a whole site is cleaner to close than a scattered set of roles, as LinkedIn's Israel closure and Oracle's team-level cuts both show. And when you size your savings, don't assume one round per cycle. Oracle ran two in six months.