For most of the software age, a growing company was easy to recognize: it hired. More users meant more engineers to build features, more salespeople to win customers, more support staff to keep them, and more managers to coordinate the expanding operation. Headcount became the visible proof that the business was scaling. Growth had a human silhouette.
Artificial intelligence is beginning to distort that familiar picture. The largest technology companies are committing extraordinary sums to AI infrastructure while becoming more cautious about the number of people they add. Meta has projected capital expenditure of up to $145 billion in 2026. Alphabet has lifted its expected spending to as much as $190 billion. Microsoft, Amazon and other hyperscale cloud companies are making similar commitments to data centers, chips, networking equipment and power supply.
Those numbers once belonged mainly to oil companies, utilities or national infrastructure programs. Their arrival at the center of the technology sector shows how much software companies now need to own before they can grow at the AI frontier. AI is not only a product feature or a subscription add-on. At the highest level of competition, companies must spend billions on compute capacity before they know whether AI will generate enough revenue to justify the cost. Growth is increasingly measured not only in people hired, but in servers, chips and data-center capacity secured.
The old internet economy was built on unusually attractive economics. Once software had been developed, serving the next million users usually cost far less than serving the first million. Talent remained the scarce input because engineers designed the product, sales teams distributed it and operations teams held the system together. In that model, hiring more people was often the practical way to turn demand into revenue.
Generative AI changes the cost structure. Advanced models depend on specialized processors, high-density data centers, cooling systems, energy contracts and vast amounts of memory. A company that wants to compete at the frontier cannot merely hire another team and ship another feature. It must secure physical capacity before demand arrives. Competitive advantage begins to depend less on software alone and more on the infrastructure needed to run it.
Layoffs and AI spending are best understood as parts of the same decision. Companies are trying to get more work done through machines and infrastructure they own, instead of through additional employees. Economists would recognize the pattern as a form of capital-labor substitution, but the speed and scale are unusual because the work being reorganized is white-collar, cognitive and often performed by people who once seemed insulated from automation.
AI does not need to replace an entire occupation to change hiring behavior. A system that drafts code, summarizes research, handles customer queries, produces sales lists or prepares first-pass legal documents may leave humans in charge while still reducing the number of humans required at the margin. The clearest early effect may not be mass unemployment. More often it is a slower refill rate. Vacant roles remain vacant. Contractor budgets shrink. Junior teams become thinner than they would have been in an earlier expansion cycle.
The evidence across the broader economy remains mixed, which makes the moment easy to overstate. The U.S. Census Bureau’s Business Trends and Outlook Survey showed AI use among American firms reaching 20.6 percent in June 2026, while Goldman Sachs analysts described the labor-market impact as visible but narrow. A Ramp and Revelio Labs study of roughly 22,000 firms found that heavy AI adopters were often increasing hiring, including at entry level.
Other signals point in the opposite direction. Goldman Sachs economist Elsie Peng estimated that AI had reduced monthly U.S. job growth by about 16,000 jobs over the previous year, with pressure concentrated in highly exposed occupations. Joseph Briggs, Goldman’s head of global economics research, has argued that sectors such as technology, consulting and graphic design are already seeing AI cut 10,000 to 15,000 jobs from monthly employment growth. The disagreement is not about whether AI is spreading. It is about who feels the pressure first: skilled workers companies still want to hire, or entry-level workers whose routine tasks are easier to automate.
Some companies have already made that narrowing explicit. In 2025, Shopify chief executive Tobi Lutke told employees that before asking for more headcount or resources, teams had to show why they could not get the work done using AI. The memo, which Lutke posted publicly after it leaked, also made AI usage part of performance and peer review. That is not a layoff announcement. It is something subtler: a new burden of proof before a job is created.
Entry-level work is where that burden becomes most consequential. Companies still need people who can judge outputs, redesign workflows, supervise systems and understand customers, but those are rarely beginner skills. They are learned through years of drafting, checking, classifying, researching, preparing and correcting. If a hiring request now begins with the question Shopify made explicit: why can AI not do this? The first roles to face scrutiny are often the same roles that teach people how professional judgment is built.
Goldman Sachs itself offers a useful example. Chief executive David Solomon has said the bank’s entry-level hiring may “contract a little” as AI changes the work mix, even while the firm continues to hire thousands of interns and graduates. The more revealing part of his argument was about training. Solomon recalled the slow manual work of his own early banking career, then asked whether a young analyst who receives instant AI-generated answers has really absorbed what is happening.
The door does not close. It narrows, and the people who enter are expected to work with machines from the beginning.
The same logic is visible on the capital side. Meta has reportedly explored a cloud-style business, internally called Meta Compute, that would sell excess AI computing capacity or access to models hosted on its infrastructure. Zuckerberg had already told shareholders that selling compute was “on the table” if the company found it had overbuilt. Empty office space usually signals overexpansion. Excess compute, by contrast, may be rentable to other firms. The asset that supports internal automation can also become an external revenue stream.
That possibility could concentrate even more power in the few companies that can afford to build AI infrastructure at massive scale. Research by Daron Acemoglu and Pascual Restrepo describes automation as a struggle between displacement effects, which remove tasks from workers, and reinstatement effects, which create new tasks in which labor remains useful. Earlier technological revolutions generated both. Factories needed operators, maintenance crews and logistics networks. The personal computer created new layers of office work. The internet produced departments devoted to digital marketing, e-commerce, cloud operations and analytics. AI will create new work too, but its most valuable infrastructure is currently controlled by firms with the balance sheets to buy chips, secure power and absorb years of depreciation.
For workers, the practical lesson follows from that task shift. Repetitive knowledge work is becoming easier to package into prompts, workflows and review queues. The work becoming more valuable is the work that requires judgment: setting objectives, interpreting outputs, understanding customers, managing risk and recognizing when a fluent answer is wrong. That favors experienced workers and technically fluent generalists. It is harder on people whose careers once began with the routine work now being automated first.
Companies face a quieter danger. A hiring rule that asks “can AI do this?” may improve discipline, but it can also hide the training value of ordinary work. Junior employees do not only produce spreadsheets, summaries or first drafts. They learn by making them, defending them and discovering what they missed. If that apprenticeship layer disappears too quickly, firms may save money on the work that teaches people how the business actually functions.
The AI labor paradox is really about where companies now think growth will come from. Software firms once grew by adding people around code. Frontier firms are now trying to grow by putting AI into every worker’s tools, every product and every workflow before they rebuild headcount around them. The next labor market will be shaped by how much room remains for humans to learn before they are expected to supervise the machines.
