AI Competition Will Turn on Model Rules, Data Control, and Power

AI can become less contestable without a single law or company decisively closing the market. The more consequential risk, as David Sacks framed it, is a convergence of gates: permission to release models, access to enterprise data, and the electricity needed to operate compute at scale.

That is why the proposed regulatory structure matters beyond safety policy. The question is whether safeguards for frontier systems can address exceptional risks without giving incumbent labs, a new regulator, or a patchwork of state rules the practical power to determine who may compete.

Demis Hassabis’s proposal, as described in the source, would create an industry-funded, federally overseen self-regulatory organization staffed by independent technical experts. Frontier labs would submit models 30 days before release for evaluation against cyber, national-security, biological, and other high-risk concerns. The benchmarks would be updated quarterly, and the body could coordinate a slowdown in development if circumstances demanded it.

Sacks supported that structure as potentially preferable to a conventional AI agency. The core reason is speed. Model capabilities and evaluation methods change too quickly, he argued, for a standing bureaucracy to write durable technical requirements and administer release approvals without creating a queue. California’s attempted AI legislation, in his telling, illustrated the problem: rules that were already poorly matched to the technology when written would be even less relevant a year later.

Financial-market SROs such as FINRA and the National Futures Association offer the relevant analogy. They allow industry participants to establish and update rules, conduct checks, and use specialized expertise, while remaining subject to government oversight. In an AI version, Sacks said, independent experts could evaluate models for cyber, biological, weapons, and manipulation risks faster than a newly created agency could.

The alternative, in his view, is a “DMV for AI”: a permission system in which every significant model release waits for review by officials who cannot keep up with the technology. He was especially wary of the “FAA for AI” model associated with Anthropic CEO Dario Amodei. The FAA’s type-certification process may make sense for aircraft, Sacks said, but it is built for timelines measured in years. He cited new aircraft certification taking five to nine years, with major modifications taking three to five years. AI models now update in months.

His support came with five conditions. First, the organization would need broad representation, including startups and open-source developers, rather than merely the largest frontier labs. Second, it should review only genuine state-of-the-art advances, not delay models below that threshold. Third, it should focus on catastrophic risks—cyber and CBRN, or chemical, biological, radiological, and nuclear misuse—not speech, disinformation, or “microaggressions.” Fourth, it should begin voluntarily and prove it works before acquiring legal force. Fifth, it should substitute for a new agency rather than become another layer of regulation.

Those limits are designed to prevent the SRO itself from becoming a capture mechanism. A body dominated by a handful of well-capitalized labs could make compliance costly, delay smaller rivals, and turn safety testing into a practical barrier against open-source development.

Chamath Palihapitiya broadly agreed that speed matters. Money will flow into political efforts to shape AI rules in ways that favor particular companies, he said. A limited framework established early may be preferable to a broad federal apparatus or an increasingly restrictive state-by-state regime. That does not mean AI would operate outside the law: commerce rules, the Department of Justice, and ordinary legal constraints would remain. The choice, as Palihapitiya saw it, is between a narrow catastrophe-prevention system and a licensing structure through which a few actors can pull up the ladder behind them.

Even the SRO model would create a new institutional foothold. FINRA reports to the SEC; an AI equivalent would need a defined relationship with the federal government. Sacks noted that deciding where it reports would be politically contested in its own right. Software has not historically had a dedicated regulator. An AI SRO could become the beginning of one.

The case for an SRO is inseparable from the fight between closed frontier providers, open-weight systems, and lower-cost international alternatives.

Sacks’s concern was that a limited safety regime could become an opening concession. Companies might accept voluntary pre-release testing, then face demands for mandatory licensing, broad approval rules, state-level obligations, and restrictions on open-source systems. In his view, a workable federal arrangement would need a clear stopping point, including preemption of conflicting state requirements.

He singled out Anthropic as the company most visibly pursuing a strategy of steadily escalating regulation. Sacks pointed to Politico reporting about Anthropic’s state-by-state effort to encourage tougher AI guardrails. His reading was that a patchwork favors companies with the capital, lawyers, compute, and government-relations capacity to navigate it. A rule passed in one state can become the baseline for a stricter rule in the next.

That argument also explains why he treats the structure of an SRO as more important than its label. If frontier labs alone define the tests, decide which models are “frontier,” and establish the costs of compliance, then a formally independent body could still constrain their competitors. Sacks repeatedly returned to open source as the critical constituency that cannot be excluded.

The economic incentive to preserve alternatives appeared in Palihapitiya’s comparison of token costs. He cited roughly $56 per million input tokens for “Fable,” around $26 for “Soul” and Claude 4.8, approximately $1 to $1.50 for Grok and Meta models, and about $0.50 for Chinese models. The figures were offered as illustrative market comparisons, but the implication was blunt: an enterprise that can route ordinary work to capable lower-cost models may face a vastly different cost structure from one locked into a premium frontier stack.

Thinking Machines’ release of Inkling represented the kind of alternative the panel wanted to preserve. The source showed Inkling described as an open-weights mixture-of-experts model with 975 billion total parameters. Sacks characterized its value proposition as deliberately distinct from maximum frontier intelligence: organizations can customize and fine-tune a less expensive open model for a particular task instead of paying for the most capable general-purpose system on every prompt.

That becomes important when AI costs reach finance teams. Ramp CEO Eric Glyman said token spending among Ramp customers had risen 21-fold over the preceding year. His company introduced token-spend management because, as he described it, CFOs often receive large bills after employees have already accumulated usage across models and tools.

Palihapitiya’s concern was that fast-growing token use could become material to earnings, not merely a minor operating-expense variance. Engineers will usually select the newest and strongest model available; they are not generally responsible for the resulting bill. CFOs are responsible for the bill and may begin imposing model-routing policies, budgets, rate limits, and requirements to show that a frontier model is necessary.

Jason Calacanis offered a practical example of the model-routing opportunity. Using Perplexity Computer with Grok 4.5, he built a custom podcast player that grouped clips across technology and business podcasts by subject. He said the project took a few hours and cost $11 in credits. The example was not a claim that every enterprise workload can move to inexpensive models. It showed that application design can often be separated from a permanent commitment to one provider: a company can choose the least expensive system that performs adequately for the task.

That flexibility is precisely what the panel feared broad regulation could narrow. If lower-cost open models remain available, the premium frontier labs will still have customers for the work that requires their capabilities. If regulation prevents users from accessing or improving alternatives, cost and compliance can become as decisive as model quality.

The source’s account of a reported Grok Build data leak made a more practical version of the control problem visible. Calacanis said the coding tool had told users that codebase data would not be transmitted to xAI servers during a session. According to the reports he summarized, the tool instead sent developers’ entire codebases to cloud servers, potentially including passwords, API keys, and change logs rather than only files relevant to a requested task.

Calacanis said the upload was disabled on July 13, Elon Musk said prior uploads had been deleted, and the Grok Build harness was later open-sourced. The episode did not turn on whether the upload was intentional. It turned on the difficulty of knowing where sensitive information travels after an AI tool gains access to a working environment.

Palihapitiya called AI privacy “fragile” and “brittle,” including when providers are trying to act responsibly. A zero-data-retention setting may reflect a vendor’s best effort, he said, but it cannot guarantee that an unknown configuration issue, software interaction, or unrecognized pathway will not expose information.

His proposed answer was a stratified ecosystem. Enterprises should not simply let employees and applications connect directly to every model provider. They need an independent layer that manages access, data handling, and exposure between proprietary systems and external models. Palihapitiya acknowledged that this aligns with the work of his company, 8090, but argued that the Grok Build incident illustrates why the separation is necessary.

Sacks connected that position to Satya Nadella’s “Reverse Information Paradox.” Kenneth Arrow’s original information paradox concerned the difficulty of selling information: a buyer cannot know its value without receiving it, at which point the information has effectively been given away. The reversal, as Sacks summarized Nadella’s argument, is that enterprises seeking AI assistance can surrender valuable internal knowledge while trying to obtain value from a model.

Nadella’s proposed trust boundary included private evaluations, proprietary learning loops within the customer tenant, decoupled orchestration, and the explicit right to fine-tune outputs. The objective is operational control over a company’s data, compute, models, and the proprietary advantage derived from them.

That approach offers an alternative to a monolithic closed stack in which one provider controls the model, harness, data pathway, and optimization loop. It may be more work for the enterprise to manage several model providers, open models, and custom fine-tuning. But it can reduce lock-in, contain exposure, and preserve the enterprise’s own “alpha.”

Apple’s lawsuit against OpenAI and several former Apple employees placed the same issue in a different setting: the boundary between an employee’s experience and a former employer’s confidential material.

The complaint shown in the source was filed in the US District Court for the Northern District of California and alleges trade-secret misappropriation and breach of contract. Apple named OpenAI, IO Products, Tang Tan—Apple’s former vice president of iPhone design and now OpenAI’s chief hardware officer—and Chang Liu, a former Apple technical engineer. The complaint alleges that job candidates were directed to bring “actual parts” to interviews and that Liu accessed Apple network storage.

Chamath Palihapitiya said Apple has not appeared especially litigious during his 25 years in Silicon Valley, making the suit a concerning signal for OpenAI. But he did not claim to know what happened and said the court would have to sort out the allegations. His principle was simpler: employees can bring their accumulated knowledge and judgment to a new job, but nobody should take materials belonging to a former employer.

Sacks stated the same rule more sharply: take only what is in your head. Skills, memories, and experience move with a worker. Thumb drives, documents, files, parts, and other proprietary materials do not.

New York Governor Kathy Hochul announced what the source described as the nation’s first statewide moratorium on hyperscale data centers. In the Reuters clip shown during the discussion, Hochul said that facilities powered by fossil fuels increase the carbon footprint, occupy land, may displace agriculture and open space, and can raise utility bills, deplete water supplies, and create noise pollution.

“Progress shouldn’t arrive with a higher utility bill, depleted water supplies, or noise pollution,” Hochul said before signing the moratorium.

Sacks rejected that account point by point. His central distinction was between facilities competing for power on an already constrained grid and facilities that build generation behind the meter. A grid-connected data center can compete with residential and commercial customers for electricity. A project that produces its own power on site changes the utility-cost argument, he said.

Behind-the-meter generation has become more appealing because standard grid connections can involve long waits. Palihapitiya described the tradeoff: a developer can build generation on its own property, but still faces complex permitting requirements, particularly for emissions. Solar requires substantial land; batteries still require power generation somewhere in the system. Natural-gas generation, including mobile turbine capacity, offers a more immediate path for operators that cannot wait for new transmission and grid capacity.

Palihapitiya described Elon Musk’s acquisition of a gas-turbine company as a move to secure direct power for future xAI data-center capacity. In his account, Musk’s use of mobile engines around the Colossus project in Memphis demonstrated how developers can pursue behind-the-meter power when conventional interconnection is constrained.

The larger issue, Palihapitiya argued, is not land availability but a shortage of electricity. He cited a PJM auction serving 13 states, including Pennsylvania, New Jersey, and Maryland, where he said expected requirements of roughly seven or eight gigawatts received only about 156 megawatts of supply. He also projected that by 2050 the United States would be short energy equivalent to 2.5 times California’s consumption.

His experience evaluating data-center assets led to the same conclusion. Sites with verifiable, energizable capacity today command unusually high prices because future capacity is uncertain. He estimated that roughly 40% of projects are being mothballed or stopped. The immediate value is not simply owning land; it is owning land where power can actually be delivered.

David Sacks disputed claims that data centers are inherently poor uses of land or water. Data centers, he argued, generate high economic output relative to their physical footprint. Noise can be managed through siting. Modern facilities can recirculate water in closed-loop systems, and he cited a study he described as finding that a typical data center used water comparable to two and a half In-N-Out Burger locations.

Palihapitiya added tax revenues, construction activity, and ongoing employment. He referred to reports of large teacher bonuses supported by data-center tax receipts in a state he recalled as North Dakota or a similar jurisdiction. The point was that a facility’s local effects depend on its power source, design, tax arrangement, and surrounding infrastructure—not merely on the fact that it contains servers.

Sacks argued that a temporary moratorium could function as a much longer interruption. Even if a state lifts restrictions in a few years, projects would still need financing, permitting, construction, and energization. He speculated that a future administration could use the lifting of a moratorium as leverage for a broader regulatory settlement. The practical result, he said, is that a one-year or two-year pause may mean many years before a new facility is running.

Calacanis condensed the dynamic into a memorable forecast: GPUs will chase energy. If major US states restrict construction while other regions make capacity available, capital, compute, and AI workloads will move toward jurisdictions that can supply power and permits.

David Friedberg argued that the political response to data centers resembles earlier campaigns against technologies whose local costs became detached from their broader scientific or economic context.

He showed a Google Trends chart comparing US search interest in “GMO” and “RT,” the Russian media outlet Russia Today, between 2010 and mid-2026. The chart was illustrative of his argument: RT entered the US market in 2010, and public opposition to GMOs increased during the period in which he believes Russian media narratives helped circulate anti-GMO sentiment through activist networks, blogs, social feeds, and mainstream discussion.

Friedberg’s concern was that data centers may be entering a similar cycle. A local project can raise legitimate questions about energy, water, land, transmission, and emissions. But a broader anti-data-center narrative can flatten those distinctions into a general hostility toward AI infrastructure. He referred to polling showing that more than half of Americans believe data centers raise water and electricity costs, including in cases where facilities recycle water or generate their own electricity.

The political mechanism he described was not limited to foreign actors. Activist organizations, media outlets, social platforms, and local opposition can all turn a complicated infrastructure question into a broad symbolic conflict about technology, wealth, and AI. But Friedberg worried that foreign interests can amplify a message when it aligns with their strategic objectives.

Sacks pointed to an OpenAI report titled “PRC-linked influence operations are targeting AI debates in the US,” along with reporting that China-linked campaigns had sought to shape US attitudes toward AI data centers. His strategic case was straightforward: China benefits if the United States limits the infrastructure required to train and deploy AI while China continues building its own capacity.

Palihapitiya connected that concern to model economics. A foreign competitor does not need to create a radically superior model if American companies become 50 to 100 times more expensive through energy constraints, closed-stack dependence, and rules that limit lower-cost alternatives. Competition can turn on cost structure as much as it turns on raw intelligence.

That is also why the panel linked data-center policy to open source. If US developers cannot build, use, or improve cheaper systems while other countries can, then the domestic market may lose its advantage through a combination of higher inference costs, restricted capacity, and reduced experimentation.

Sacks argued that the current AI debate has become a moral panic around harms that have not occurred at the scale implied by the most alarmed rhetoric. He cited fears about catastrophic cyber misuse and widespread job loss. Calacanis took a narrower position: policymakers and companies should monitor cybersecurity, labor displacement, and autonomous systems; frontier labs should continue red-teaming and safety testing. Monitoring is not the same as panic.

Sacks warned that a narrowing lead over China makes delay more consequential. He mentioned Kimi K3 as a model that people were describing as close to the frontier. The practical concern was not whether a particular benchmark claim endures. It was that the United States may be erecting costly approval systems and limiting infrastructure just as Chinese models become more competitive.

David Friedberg closed with a concrete application of AI that bears little resemblance to the policy fight over regulation and infrastructure: using computational protein design to develop a new biological tool.

The reported laboratory result concerned carboxymethyllysine, or CML, an advanced glycation end product that can accumulate in the extracellular matrix—the material between cells. Friedberg described glycation as the process by which sugars and fats bind to proteins over time. As those altered proteins accumulate, they can become harder for the body to clear and repair; collagen and other structural proteins can change shape, bind together, lose flexibility, and contribute to inflammation.

Researchers at Calico and Revel Pharmaceuticals sought an enzyme capable of degrading CML. Friedberg said they used AlphaFold to identify a protein that could bind to the target, then used directed evolution to improve it. That meant altering DNA sequences, producing many protein variants, testing their activity, and repeating the process through five cycles.

The result, as Friedberg described it, was an enzyme that removed 52% to 97% of CML from tested proteins including casein, collagen, retinal proteins, and hemoglobin. In donated skin from people older than 70, he said treatment eliminated roughly 55% of CML, bringing the skin’s measured age to approximately that of a 31-year-old.

The delivery constraint remains central. An enzyme that performs well in a test tube or on donated tissue must still reach the relevant extracellular matrix safely and effectively in a living body. Friedberg named topical creams, injections, supplements, and RNA-based delivery as possible routes, but treated each as an open question.

Palihapitiya predicted that cosmetic skin care would likely be the first major market if topical delivery works. A cream that meaningfully reduced visible glycation-related skin aging, he said, could be commercially enormous.

The scientific claim Friedberg emphasized was not that a consumer anti-aging treatment is imminent. It was that AI-assisted protein design, combined with directed evolution and high-throughput testing, can help create proteins that do not exist in nature and perform a desired biological function. That is the kind of practical payoff he believed gets obscured when AI is discussed only as a source of social risk, regulatory concern, or enterprise expense.