A handful of data points landed in the same few weeks, and read together they describe one thing rather than several. The AI boom is not spreading out. It is concentrating. Capital, compute, and now the shape of hiring itself are all pooling around a small number of very large players. If you want a single frame for what is happening in the business of AI right now, concentration is the one that fits the evidence.
The buildout is real, and bigger than the skeptics think
There is a steady drumbeat of doubt about whether the physical infrastructure behind AI can actually get built. Power, labor, permits, supply chains, capital: every one of them is a genuine bottleneck, and some forecasts lean on that to argue the whole thing stalls. SemiAnalysis has spent months pushing against that read, and the latest data gives them a strong hand.
The firm projects roughly 50GW of new datacenter capacity delivered globally in 2027, which is close to the entire cumulative datacenter market built through 2024 arriving in a single year. That is not a hunch. It rests on tracking construction building by building, using satellite imagery and leasing data, and the pipeline is visible: tens of gigawatts already under vertical construction, much of it more than half finished. Brad Gerstner of Altimeter publicly called the forecast too aggressive and put his own number closer to 25GW. The disagreement is worth watching, because it is really a disagreement about whether the constraints bite before the concrete sets.
The more interesting wrinkle is where the real shortage sits. SemiAnalysis argues the binding constraint is shifting away from power and toward chips, with front-end semiconductor manufacturing capacity becoming the ceiling. Memory is eating a growing slice of hyperscaler capex. For investors, that moves the pressure point upstream, away from the developers laying slabs and toward the few firms that can actually fabricate the silicon. Another form of concentration, in other words.
Four out of five venture dollars
The funding picture tells the same story with cleaner numbers. Through the first three quarters of 2026, AI companies captured about 83 percent of all US venture deal value, the highest share PitchBook has recorded, out of a record market of roughly $516 billion. For comparison, internet companies took around half of venture capital at the dot-com peak. We are well past that.
Most of the money is pooling at the very top. OpenAI, Anthropic, xAI, and Waymo together accounted for close to two-thirds of global venture funding in the first half of the year. Rounds of $100 million or more have come to dominate the dollars, while smaller deals have shrunk to a sliver of the total. If you are building anything that is not an AI company, the arithmetic has turned quietly against you: a non-AI startup raising a Series A this year lands a median pre-money valuation around $42 million, against roughly $78 million for an AI company at the same stage.
That gap is where I would urge some caution. The headline valuations belong to a handful of outliers, and AI's share of US dollars has actually drifted down across the quarters of 2026 even as the annual figure looks enormous. A founder pricing a round off the 95th-percentile comparable is borrowing a number that was never meant for them. The sensible move is to price off the median and treat anything above it as upside, which is roughly the view The AI Corner reached, and I think it holds.
What agents do to a payroll
The most revealing data came from inside Anthropic. On October 16 the company published a study of its own workforce, and it is the first time a frontier lab has shown, with numbers attached, what heavy AI use does to hiring. By August, the agents its employees directed were logging around six hours of work for every human hour. Code commits per engineer were roughly five times their level six months earlier.
The number that matters for the labor market is this: the junior share of technical hires fell from 27 percent to 9 percent since early 2025, even as the company tripled in size. Normally a firm growing that fast hoovers up entry-level people. Anthropic grew without doing that. Non-technical departments, notably, held roughly steady, so this is specific to the work agents can now absorb.
Two honest caveats. The six-to-one figure and the hiring figure use different rulers, and neither directly counts jobs displaced. And the self-interest is obvious: the company selling the agents is the one reporting how transformative they are. Still, the pattern shows up in outside data too, and it points at the same concentration dynamic. The rung being pulled up is the first one.
So what
The through-line is that AI is rewarding scale and punishing the middle. Compute concentrates in a few campuses and fewer fabs, capital concentrates in a few labs, and output per worker rises in a way that thins entry-level roles. For investors, the opportunity and the fragility live in the same place, since so much rests on so few names. For founders and workers, the useful response is unglamorous: price to the median, not the outlier, and assume the bottom rung of the ladder is the one being redesigned first.