The Frontier Reopens: AI's Shift From Smarter Models to Smarter Systems

The headline version of the AI story is a single number going up: whichever lab has the smartest model wins. Listen instead to the people building and funding the frontier right now, and a more useful picture emerges. The interesting contest is no longer only about raw model intelligence. It is about how intelligence gets built, served, powered, paid for and governed. The frontier has reopened, and it is being contested on terrain that matters far more to anyone deploying this technology than another point on a benchmark.

We drew this from four recent long-form interviews: with Alexandr Wang, who leads Meta Superintelligence Labs; with Mira Murati, the former OpenAI CTO now building Thinking Machines Lab; with Andrew Feldman, CEO of the wafer-scale chip company Cerebras; and, from the capital side, with the activist investor Bill Ackman of Pershing Square. They sit in different layers of the stack and disagree on plenty. What is striking is where they converge.

The Frontier Is Wide Open Again

A year ago, the conventional wisdom held that the frontier had consolidated around a handful of labs and that everyone else was permanently behind. That story is unravelling. By its own AI chief's account, Meta concluded that its earlier models were not on the trajectory it needed, stood up a new lab, rebuilt its scaling ladder from scratch, and shipped a deliberately modest first model it cheerfully calls an appetizer, with the entree still cooking. An entirely new lab, meanwhile, has raised serious capital to pursue a different architecture of human-AI interaction, and a specialist chipmaker most investors had written off is suddenly indispensable.

The builders frame this openness as healthy. Advancing the frontier, Murati argues, is highly positive-sum: there is room for a plurality of approaches, and the barrier to entry is high enough that the few who clear it can still differentiate. Feldman puts the same idea in the language of evolutionary biology. Whether a specialist beats a generalist, he says, depends entirely on the shape of the resource landscape. When demand concentrates into one enormous vein, as AI compute now has, purpose-built specialists win; when it fragments into many small pockets, generalists do. From the investor's chair, Ackman sees the same dynamism, and treats it as the defining risk of the era.

This is the greatest era in history to build a business, and that means the probability of being disrupted has gone up enormously.

Unlimited access to compute, capital and talent, he notes, makes it far easier for two people in a garage to upend an incumbent. The hardest and most important job of a long-term investor is therefore to judge what can be disrupted and what cannot. The lesson for enterprises is the mirror image of the one for investors: do not bet on a single eventual winner, and do not assume your own position is safe. Capability, and competition, will keep arriving from multiple shifting directions.

Intelligence Is Not Enough: Speed Is the New Battleground

For two years the race was about training ever-larger models. The builders are now preoccupied with the other half of the problem: serving them. Once AI became smart enough to use daily, Feldman points out, speed started to matter the way it always eventually does for any technology that embeds itself in everyday work. He reaches for the obvious analogy, that nobody would willingly trade broadband for dial-up, and a less obvious data point: even single-digit-millisecond delays measurably reduce how much people use a service. Nobody wants slow AI.

That conviction is reshaping system design. Cerebras splits inference into its two very different jobs, processing the prompt (which is parallelisable) and generating the answer token by token (which is strictly sequential), and runs each on the hardware best suited to it, even pairing its own chips with a hyperscaler's for the other half. It is a concrete example of a broader shift: the advantage increasingly comes from architecting the whole serving pipeline rather than from owning the single cleverest model. Wang's emphasis on agentic, multimodal capability and Murati's real-time interaction models point the same way. Ackman adds the financial corollary, observing that time itself has become more valuable in the AI era, where losing even a couple of months can be decisive. The frontier is moving from what a model knows to how quickly and fluidly it can act.

The Real Bottleneck Is Physical

Ask the people building this what actually constrains them, and the answer is not algorithms. It is data centers: power, land, cooling and fiber. Every serious player, Feldman notes, is gated by the deployment of physical infrastructure, not by ideas. And against the constant talk of an AI bubble, he offers a contrarian read. Classic bubbles are defined by speculative overbuilding, the way fiber was over-laid in the late 1990s on the assumption that demand would follow. This moment, he argues, is the inverse.

What is unusual about AI right now is that the builders are so far behind the demand it's absurd.

Customers move at the speed of software, he says; builders move at the speed of real estate. That gap, with backlogs measured in tens of billions, is the opposite of overbuilding. It also comes with a self-inflicted cost: community backlash against data centers. Here Feldman is candid that the industry behaved like amateurs, racing ahead instead of being a good neighbor. The facts, he argues, are favorable when anyone bothers to share them. A mid-sized data center creates thousands of construction jobs and broadens the local tax base, and the entire U.S. data-center fleet consumes several times less water than California's almond growers. The failure was one of trust and communication, not physics.

Ackman supplies the counterweight from the capital side. Underwriting the leading AI companies, he says, looks less like buying a stable business and more like late-stage venture: you weigh the people, the opportunity, the context and the deal. The harder part is that several of these companies are making capital commitments far in excess of their revenues, which he calls a genuinely difficult thing to underwrite. Demand running ahead of supply and balance sheets running ahead of revenue are both true at once. The discipline is to hold both facts in view rather than choosing the one that flatters your position.

The Economics Grow Up

If 2025 was the year of treating tokens as if they were free, the builders now see the discipline phase arriving. Feldman's analogy is shopping at a warehouse store for the first time: you start by walking every aisle and going home with an absurd tub of mayonnaise, and only later learn to shop deliberately. Organisations are learning the same lesson with tokens, realising that you do not reach for the most expensive model for every task.

You don't need a Ferrari to go to the grocery store.

The mature pattern is allocation: route routine work to cheaper or open-source models, reserve frontier models for problems that genuinely need them, and meter usage the way any organisation rations a real resource. The harder truth sits one layer down, in adoption. Ackman, who knows most of the S&P 500 either directly or one call away, says AI is now the first question in nearly every boardroom, ranked simultaneously as the biggest opportunity and the biggest threat. Yet real success remains rare. He points to research suggesting the large majority of enterprise AI initiatives fail, and notes that the hottest job in the field is the forward-deployed engineer whose entire purpose is to close the gap between AI's promise and its return. By his own admission even his firm is early, with its first durable use cases in legal, compliance and back-office work.

The conclusion converges with what the broader market is already teaching: the durable question is shifting from how capable a system is to what return its spend produces. Builders who can answer that with evidence will keep their budgets. Those who cannot will watch them contract as scrutiny arrives.

Keep Humans in the Loop, by Design

The deepest disagreement in the industry is not about capability but about the role humans should play as capability grows. One path advances AI as autonomously as possible, treating the messiness of human interaction as friction to be removed. Murati's entire bet runs the other way. Her lab's interaction models are continuous rather than turn-based, taking in audio, text and video in small slices and responding in real time, capturing the interruptions, pauses and overlaps that carry so much meaning in human exchange. The goal is to build AI as a genuine tool for thought that keeps people in the loop.

It's more like a tandem bike, where both people are pedalling and both hands are on the wheel.

This is a design philosophy, not a sentiment. Murati frames keeping humans in the loop not as a sign-off checkpoint but as a system architecture that produces both usefulness and alignment, and she argues the discipline matters most now, while the discontinuities between capability jumps are still small. Wang's vision rhymes with it from the consumer side: not an app for every task, but one or two trusted agents per person, a personal agent woven into health and relationships, perhaps a separate one for work, that people come to rely on the way they rely on email. Either way, the message for builders is that human in the loop is an architectural choice with consequences, not a compliance box.

Trust, Safety and Who Gets to Decide

As the models grow more capable, all four voices circle the same theme from different angles: capability is outrunning the institutions meant to govern it. Wang is candid that Meta declined to open-source its newest model because early testing triggered elevated risks, including in biosecurity, that are far harder to mitigate once weights are released into contexts no one controls. The company is now working on versions safe to open-source without giving up too much capability. He treats the trustworthiness of personal agents as one of the defining societal questions of the agentic era, and as an industry-wide problem rather than any one company's.

Murati pushes the point further. The character of the people building these systems matters, she says, but the conversation fixates there and neglects something more important: institutional design. Well-intentioned people still make mistakes and misjudge consequences.

Morality is not everything. You have to think about decision-making structures, transparency and governance.

Her conclusion is that power should not hinge on any single person; it should rest on checks and balances, distributed agency and shared knowledge, so that more people can weigh decisions for themselves. Even Feldman's plea for the industry to behave like a good neighbor is, at bottom, an argument about trust. The common thread is that governance, institutional, technical and social, is no longer a side conversation. It is becoming part of the product.

What This Means for Builders

Strip away the personalities and the rivalries, and these four vantage points converge on a practical agenda for anyone deploying AI in earnest:

• Do not bet on a single winner. Capability and disruption both arrive from multiple shifting sources; design for optionality across models and providers, and assume your own position is contestable.
• Engineer for speed, not just intelligence. Latency shapes adoption; architect the whole serving pipeline, and match each job to the hardware and model that fit it.
• Treat compute and power as strategy. The binding near-term constraints are data centers, electrons and permits; plan capacity, and community goodwill, accordingly.
• Allocate spend like a scarce resource. Route routine work to cheaper or open models, reserve frontier calls for what needs them, and measure return rather than activity.
• Close the ROI gap deliberately. Most enterprise initiatives stall in the space between promise and production; invest in the people and the plumbing that turn capability into outcomes.
• Make human in the loop an architecture. Build for genuine collaboration and oversight, not a checkpoint, because that is where usefulness and alignment both come from.
• Treat governance as product. Trust, safety and clear decision-making structures are now features that users and regulators will judge you on.

None of this is as exciting as a leaderboard. But it is where the people building and funding the frontier are actually spending their time, and it is a far better guide to what will work in production than the next record-breaking model. The race for raw intelligence will keep generating headlines. The race to build intelligent systems that are fast, affordable, well-powered, human-centred and trustworthy is the one that will decide who actually benefits.