AI’s Next Constraint Is Turning Capability Into Use

Google’s product challenge is not simply to make Gemini smarter. It is to make AI work across the services people already use, rather than leaving them to navigate separate tools. Sundar Pichai argues that frontier research matters because it helps Google see what is coming and improve future models faster. But Margins’ Ranjan Roy sees a different bottleneck: Google’s AI products still feel disconnected, including Gemini in Workspace and Gemini as a standalone product.

The distinction matters because the proposed solution is not just a better feature inside Gmail. Pichai described a unified agent, initially for enterprise and later for consumers, that people could reach through Workspace, Search or Gemini. Big Technology’s Alex Kantrowitz sees that as a change in the product’s center: the agent becomes the interface, with Google’s existing services working behind it. Roy agrees with the direction, but emphasizes what it would demand inside the company. Product teams accustomed to measuring daily users or engagement would have to treat their services as components of an experience that may begin somewhere else.

The disagreement is over what makes that agent useful. Roy gives more weight to product execution: an agent should respond well to ordinary requests, avoid overcomplicating simple tasks and guide users through work. Kantrowitz argues that the model’s intelligence matters too. A useful agent must understand what a person intends, choose how much effort a task warrants and know when to interrupt. Neither argument reduces to “interface versus model.” The practical question is how capability gets translated into an experience that works across services.

Genie 4 offers a contrasting example of product design: its demonstrations center on continued direction, rather than a single prompt followed by a finished output. A user can move through a generated scene, then change what happens with further instructions. In one example, a crane walks through a snowy landscape; the user adds a shooting star and then directs the crane toward a piano. DeepMind presents this as an environment that can be steered as it unfolds.

That interaction is different from asking an agent to carry out a task across email or spreadsheets, but it sharpens the same question: how much should the user direct, and how much should the system decide? DeepMind proposes virtual robot practice as a possible use for Genie 4, showing a simulated arm performing a laboratory task. That is a proposed application, not evidence that the system is ready to train robots for physical work. Access is also limited to a research preview and selected testers and partners. The demonstration makes a new interaction legible; it does not establish how broadly useful the system will be.

Google faces a more difficult version of the same product problem. It has broad distribution and many services to connect, but its product groups may have different priorities and measures of success. Genie 4’s demonstrations suggest one way to design for user direction. Google’s agent proposal raises the organizational question behind it: can a company make its capabilities feel like one coherent product?

Google’s AI Advantage Depends on Coordinating Products, Compute, and ResearchAlex Kantrowitz

Genie 4 Lets Users Steer Generated Worlds as They UnfoldGoogle DeepMind

A wave of expected GPU supply has not meant that compute is cheap or readily available. Investor Gavin Baker said rental prices for nearly every kind of GPU rose through October, with tightness extending from current capacity into future contracts. He described companies struggling to secure compute well into 2027 even as substantial new supply is expected.

The distinction is between announced capacity and working capacity. Operators need financing, equipment, power, and the ability to bring clusters online and run them reliably. Baker said some firms can take three to six months to energize GPUs. Equipment that is bought but not yet operating is costly capital that cannot generate revenue. His comparison to running a Formula One team is pointed: buying the hardware is only the beginning.

The economics also depend on how operators sell future capacity. Baker favors more exposure to spot prices and customer prepayments over long-term fixed-price commitments, arguing that fixed deals can leave operators selling compute below market when prices rise. That is his investment interpretation, not a settled prescription for every operator. The same uncertainty applies to his forecasts about future prices, investment returns and whether infrastructure firms are nearing the peak of external financing.

Google’s compute choices show that scarcity is not only a market problem. Pichai said Google Cloud was giving up short-term revenue to support frontier AI research. The company is deciding whether scarce capacity should serve customers who can pay for it now or researchers whose work may improve future models. Google Cloud’s reported growth makes the trade-off harder to dismiss as theoretical.

Kantrowitz argues that continued growth could put Google in contention to pass Azure, strengthening the case for allocating more capacity to Cloud. Roy would give Cloud more, on the grounds that most users do not need a frontier model for most tasks. Pichai, by contrast, has described frontier research as important to Google’s ability to improve models. These positions expose the opportunity cost: capacity committed to research cannot simultaneously be sold to Cloud customers.

Baker’s account complicates any assumption that new supply will quickly relieve the constraint. Easier-to-deploy systems may reduce delays, and more equipment is expected. But whether that supply becomes useful compute depends on execution. In Google’s case, it also depends on which internal priorities receive the capacity first.

GPU Prices Are Rising Despite a Wave of New Compute SupplyTBPN

In law, the case for AI begins with a shortage of capacity rather than a shortage of tasks. Legora chief executive Max Junestrand argues that demand for legal services exceeds the time and staffing available to meet it. Contracts may go unreviewed; diligence may happen only after an investment decision; some disputes may never proceed. AI could expand what legal teams can take on, including by handling a first pass at contract review or more of the work in a transaction.

The useful distinction is between doing existing work faster and making previously neglected work possible. Junestrand’s examples point to both. If a team has more capacity, it might review more contracts or do diligence earlier, rather than simply shorten the turnaround on the same workload. But the system’s ability to produce work is not enough. Junestrand says lawyers must be able to trust what it produces and must not feel they are second-guessing it.

Legora cites results from one Fortune 100 customer: a 90% reduction in contract turnaround time, more than $20 million saved on external litigation and diligence, and 9.5 hours returned to each lawyer per week for higher-value work. Those are company-reported results for one customer, not proof of equivalent gains across the legal profession. They illustrate the scale of the claim while leaving the central adoption question intact: will lawyers rely on the work enough to change how they practice?

That question echoes Google’s integration problem. A capable model or agent is not useful in a professional setting simply because it can perform a task. It has to fit the workflow, produce work people can assess and earn enough trust to be used without constant rechecking. Junestrand also describes Legora’s product as something that must be rebuilt as AI capabilities advance, rather than treated as finished. The evolving model is an input; the product and the institution must keep adapting around it.

The comparison is not that legal software and a unified consumer agent face identical demands. It is that both expose the distance between capability and use. In law, the stakes of relying on a system make trust explicit. In Google’s case, the challenge is to integrate services and reorganize product ownership so that a user can rely on an agent across them. In each, adoption depends on a change to the work around the model.

AI Could Expand Legal Capacity Beyond Current LimitsOpenAI

Alpha School principal Joe Liemandt’s proposal starts from a different kind of capacity: time in the school day. He argues that if AI helps students learn core academics in less time, schools should use the hours they free for work conventional classrooms often leave to extracurriculars—teamwork, communication, agency and entrepreneurship. At Alpha, he says, students spend two hours a day on academics and the rest on projects intended to build those skills.

That is an institutional model, not settled evidence that AI will let schools everywhere compress instruction or that Alpha’s approach will generalize. Liemandt describes Alpha’s learning engine as capable of teaching ten times faster than a traditional classroom; separately, he says Alpha students learn twice as much in two academic hours as in six hours of class plus homework. Those are his claims about the model and its results. The larger point is that the school day has to be redesigned if the time savings are to matter.

Alpha’s projects are meant to make interpersonal skills part of the curriculum rather than an incidental benefit of group work. Liemandt describes students practicing teamwork through structured exercises, with AI used to analyze how they communicate. In one example, third graders complete an escape room, take an eight-week workshop and return to the task. He says they must raise the share of team-oriented language from less than 20% to 85% to pass the class. The method and thresholds are Alpha’s; they show how the school tries to define and assess a skill that is often left vague.

Liemandt and Gagan Biyani also argue that students should learn by building things with purpose, not only by completing academic assignments. Their examples include students working on software and educational projects, and Alpha’s Founders School, which treats starting a business as a skill to practice. The claims about outcomes belong to the school’s leaders, and the approach raises questions about how such programs can be evaluated and reproduced. Still, the institutional choice is clear: if AI makes some learning more efficient, does the saved time return to the same schedule, or become space for other forms of learning?

That question links the school to the workplace examples. A legal team can use added capacity to take on more work, but only if trust and workflow change. Google can build a powerful model, but a unified agent requires product teams to coordinate and accept new measures of success. In each case, the technology may create room to do more. Whether that room becomes useful depends on what the organization decides to do with it.

AI-Accelerated Academics Make Room for Real-World Skillsa16z