Self-Improving AI Is Real. The Scarcity Now Is Physical.

The most consequential AI prediction this week didn't come from a lab. It came from a conference stage in Hong Kong on August 29, where Kaifu Zhang — a former Alibaba Group vice president who headed search, recommendation, and generative AI for the Taobao and Tmall division — told investors that AI has entered a "self-improvement" stage and that the next one to two years could bring another round of exponential growth.

If that sounds like hype, hold on. Zhang now runs an AI startup, so he has every incentive to sell a story. But the claim is not actually the investment signal. The signal is what the claim does to the market's scarce inputs — and the data on that is already on the books.

Here is what "self-improvement" literally means, stripped of the sci-fi. Not an AI awakening. Something narrower and more mechanical: AI-written code merges itself into production. AI generates the data used to train the next generation of AI. AI runs the experiments that optimize the models. Close that loop and the development cycle stops depending on human hands — and human hands stop being the bottleneck.

The evidence that this loop is real comes from the lab that has the most to gain from you believing it, so read it as self-reported but read it anyway. Anthropic reports that, as of May 2026, Claude has authored more than 80% of the code merged into Anthropic's own codebase, up from low single digits in early 2025. The length of task a model can complete on its own has been roughly doubling every four months: four-minute tasks in March 2024, twelve-hour tasks by March 2026. That is the mechanism in two numbers — machines writing code that builds the machines that do the next quarter's work.

Now the honest counterargument, because it is real and it matters. In August 2026, MIT Technology Review covered a Princeton-led test that handed Anthropic's most capable model a genuinely hard assignment: produce original research good enough for a top machine-learning conference. The original paper authors rejected both papers the model produced. The researchers found the AI superb at "research engineering" — running hundreds of experiments, compiling results — and "unambiguously bad" at the research itself. No novel contributions. No ability to start over. Even Anthropic cofounder Jack Clark has called current AI creativity "a bearish signal on short recursive self-improvement timelines."

So grade the evidence honestly. The part of self-improvement that is measurable and real is the boring part: coding, optimization, post-training. The part that would trigger an actual intelligence explosion — open-ended research judgment, taste, knowing which problem is worth solving — is unproven. Even Zhang hedged his own headline, framing the exponential jump as conditional on AI reaching top-researcher-level optimization efficiency.

But here is the twist that matters for money. Both versions of the story point the same tick at the capex. Zhang's own description of the transition is the key: agents shift from "passive response" to "always-on, active task execution". Translation: AI stops being a tool you query and becomes a worker you leave running. A worker you leave running burns money continuously.

The numbers confirm the shift is already happening. Industry estimates put inference — the cost of running models, as opposed to the one-time cost of training them — at roughly two-thirds of all AI compute in 2026. NVIDIA told its own conference audience at GTC that inference demand has grown around a millionfold in two years, driven by exactly these always-on, agentic workloads. And the new workloads are not cheap to run: reasoning and multi-step agent tasks consume hundreds to thousands of times more energy per query than a simple text answer.

This is where the abundance-scarcity frame does its real work. If AI improves itself, intelligence — the scarce input of 2023, gated by human researchers and human data — becomes abundant and cheap. The scarce complements turn physical. And the market has already voted with its checkbook. The IEA reports that hyperscaler capital spending exceeded $400 billion in 2025 and is budgeted to jump another 75% in 2026. Five technology companies now spend more capital than the entire world invests in oil and natural gas production. Data center electricity is projected to roughly double from 485 terawatt-hours in 2025 to 950 in 2030 — about 3% of the world's electricity. High-bandwidth memory, the physical component shortest on supply, is sold through at least 2027, by the agency's account.

Stop and sit with that oil-and-gas inversion for a second, because it is the single most underrated fact in this story. For a century, oil companies were the marginal buyers that set the price of the world's most important energy. That buyer has changed. The marginal buyer of the world's new electricity is now a handful of AI companies. That rewires pricing power across the entire energy economy — and it puts a standing bid under gigawatts the way the 2020s put a standing bid under silicon.

So what does a thoughtful investor actually do with "another exponential jump in one to two years"? The asymmetry is the point. You do not have to believe Zhang's timing — or even his central claim — for the direction of money to have been set. The $400-plus billion was spent. The 75% increase is budgeted. The power is being contracted. Capex is a now-event; the exponential is a maybe-event. When a forecast and a cash flow point the same direction, the cash flow is the evidence.

Two cautions, stated plainly. First, this is a crowded trade. The semiconductor ETF SMH has absorbed more than $8.5 billion of net new money this year, and the boom is priced into the obvious names — which is exactly where a beginner's money goes first. The durable edge, if it exists, sits in the parts of the scarce layer the market prices later: electricity and the grid, memory, and the identity-and-security software that must exist before an always-on agent is allowed to touch real systems — Zhang's own list of what the transition demands.

Second, and this is the invalidation to watch: efficiency, not hype. The IEA notes that energy use per AI task is already falling by an order of magnitude every year. If efficiency gains outrun usage growth, the inference boom stalls, the capex cycle cools, and the scarcity story inverts. Watch the power curve and the capex guidance — not the conference quotes.

The framework, in one line: when a technology makes something abundant, durable value migrates to what it makes scarce. If AI is about to improve itself, intelligence stops being the scarce input. Gigawatts are. Memory is. Trust is. The man who ran Alibaba's AI gave the thesis. The capex is the proof. The power meter is the verdict.