AI Economics Are Splitting Between Compute, Distribution, and Frontier Models

The consequential divide in AI is not simply between proprietary and open models. It is between businesses with visible near-term returns—selling compute, distributing AI through established products, or operating mature software—and bets whose returns depend on retaining a technological lead that is expensive to build and difficult to prove durable.

David Friedberg reads Google’s AI reorganization primarily as a capital-allocation decision, rather than as evidence that the company has simply lost control of its research organization. Alphabet projected $195 billion to $205 billion in 2026 capital expenditures, much of it directed to AI infrastructure and data centers. Friedberg argues that this spending has a comparatively legible return: demand for compute is intense, Google knows how to run global infrastructure, and accelerated depreciation makes domestic compute spending tax-advantaged. At a 26% corporate tax rate, he said, immediate expensing effectively returns 26 cents on every qualifying dollar of capex in the year it is deployed.

Frontier-model development has a different profile. It can require tens of billions of dollars, while the path from a leading model to enduring profits is less clear as open-weight systems improve and rival labs narrow the gap. Friedberg’s distinction is that infrastructure capex is “high alpha, low beta,” whereas frontier-model development may have substantial upside but remains a much riskier use of capital.

That framework explains his interpretation of the changes at Google. Demis Hassabis moved from chief executive of Google DeepMind to chairman and chief scientist, while Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals left to form Discovery Loop, a company focused on scientific discovery through AI. If Google increasingly sees its comparative advantage in supplying the larger ecosystem rather than putting its largest internal bet behind one general-purpose frontier model, researchers whose ambitions center on scientific breakthroughs may have a reason to leave.

Brad Gerstner describes the tension as a “channel conflict.” Google Cloud wants compute that it can rent to external customers, including AI labs that compete with Google’s own model teams. Those internal teams want the same compute to build systems that compete with those customers. Gerstner sees comparable tensions at Microsoft and SpaceX: infrastructure offers more visible returns, while a frontier-model program consumes scarce resources with a less certain payoff.

Google’s reported results support the infrastructure case without resolving the question of frontier leadership. The company said Cloud revenue grew 82% year over year, driven by AI infrastructure and AI solutions, and that Cloud backlog reached $614 billion. It also said nearly 90% of the Fortune 500 uses Gemini Enterprise. Jason Calacanis emphasized Google’s consumer distribution: Android, Search, Gmail, Chrome, and YouTube each have more than 3 billion monthly users, and Gemini had more than 950 million monthly active users in the second quarter, tripling year over year.

Friedberg’s broader point is that Google need not own the best model in every category to build a major AI business. It can distribute Gemini through existing products, host open-weight models, work with outside labs, and compete through specialized systems in areas such as video, genomics, life sciences, and protein folding. He noted that Hassabis remains involved with Isomorphic Labs.

The dispute is over whether frontier intelligence itself will retain a premium. David Sacks argues that the market is bifurcating. At the top sits frontier intelligence, where he sees Anthropic and OpenAI as the leading pure plays. Below it is a large market for models six to twelve months behind the frontier: useful and broadly deployable, but unable to command significant payment for the weights themselves.

Companies outside the frontier can still charge for compute, inference, implementation, and consulting, Sacks says. But they cannot charge a major premium for the model layer. He compares the structure to Apple and Android: an open ecosystem can have broad usage, while the premium product captures disproportionate monetization.

Sacks cited Anthropic’s reported growth as evidence that frontier performance has not become a commodity. He said Anthropic had surpassed $80 billion in annual recurring revenue after starting the year around $10 billion, with estimates rising above an earlier $100 billion year-end target. His claim is not that every task requires the best available model. It is that competitive businesses and companies still discovering AI use cases have reason to pay for it. A hedge fund may not accept even a small risk that a cheaper model leaves it behind rivals; a company developing an unfamiliar workflow may not yet understand its input distribution, failure modes, or quality thresholds.

Calacanis disputes how wide that premium market really is. He said the difference between the open-source models he uses and frontier systems is already negligible for his ordinary work. Elon Musk publicly replied that the difference was “actually a world of difference.” Calacanis’s view is that most users who do not need unusual speed or demanding technical performance can already use open models, and that deployment will only become easier.

The Decagon post shown during the discussion supplies a maturity curve that accommodates both positions. New enterprise applications start with frontier models because the company needs broad intelligence while it is discovering the problem. Once a workflow is well understood—its expected inputs, desired outputs, failure modes, and quality bar—a company can distill, fine-tune, and specialize smaller open models. Decagon presented customer service as an obvious example: a mature, high-volume workflow where a smaller, faster model can be shaped around a defined task.

Friedberg expects enterprises to use a portfolio rather than choose one permanent winner. Simple workflows may run on inexpensive open-weight systems; quality-critical work may justify a premium model; specialized tasks may use the best available video or life-sciences model. Consumers, meanwhile, may be content to pay $20 or $40 a month for a managed service such as ChatGPT, Gemini, or Claude.

Gerstner adds that open models are not automatically cheaper after operating costs. He cited Jensen Huang’s argument that a company choosing to operate its own model must bear the costs of training, fine-tuning, maintenance, guardrails, evaluation, hosting, and safety. Open systems matter when customers need control and specialized adaptation, but the absence of a model-license fee does not make the full system cheap.

For Gerstner, that is the desirable outcome of competition: open models put pressure on pricing, while frontier labs retain premium economics where performance matters. He agrees that Anthropic and OpenAI are currently the clearest frontier pure plays, but resists treating the market as a settled duopoly while Google, Microsoft, Amazon, Meta, and other well-capitalized companies remain active.

The SpaceX and Airtable cases point to opposite versions of the same problem. SpaceX has an operating business that may generate cash to fund large new bets, but its compute expansion depends on scarcity pricing and a vast financing requirement. Airtable has revenue, cash, and customers, but its venture-capital structure no longer appears suited to a slower-growth software business.

SpaceX reported $7.8 billion in second-quarter revenue, up 92% year over year, and $3.5 billion in adjusted EBITDA, up 119%. It reported $100 billion in cash, cash equivalents, and marketable securities, plus $43.5 billion in trailing-12-month free cash flow. The company’s results document listed 12 million Starlink subscribers, 1.4 gigawatts of deployed AI nameplate capacity, and 78 launches year to date.

The hardest question is not whether compute can earn revenue at today’s prices. It is whether SpaceX can build enough capacity, finance it, and retain high rental rates long enough for those investments to earn an adequate return.

SpaceX expected about 2 gigawatts of compute by year-end. Sacks said Musk indicated a range of 5 to 10 gigawatts in the following year, closer to 10. Gerstner used a minimum build cost of $50 billion per gigawatt. On Sacks’s illustrative assumption of moving from 2 to 8 gigawatts, the incremental buildout would require roughly $300 billion of capital.

David Sacks offered the optimistic arithmetic: at 2 gigawatts rented for $50 per watt, compute alone could imply $100 billion in annual recurring revenue. That is an estimate based on the spot-pricing assumption, not a reported SpaceX forecast or result. It excludes Starlink, launch, Grok, and Cursor.

Gerstner’s concern is that the apparent one-year payback implied by present rental rates is a scarcity outcome, not a stable law of the business. He said expected payback was closer to four or five years only months earlier. Frontier labs are willing to pay exceptional prices for at-scale compute because they believe access can confer an advantage in model development. Anthropic’s large compute purchase was central to this logic, and Gerstner believes OpenAI and Anthropic would buy more capacity if it were available.

But that demand is concentrated. The ultimate offtake commitments are heavily associated with Anthropic, OpenAI, and Nvidia, while hyperscalers are building capacity to serve customers of that scale. If spot prices fall, the physical assets do not become cheaper after they are built; the return profile changes and the payback period stretches.

Brad Gerstner identified the financing choices directly: debt, dilutive equity, or supplier support such as Nvidia-backed financing. Each has costs. Nvidia shareholders may resist unlimited customer backstopping if the concern is that compute prices could turn against the buyer. And construction requires more than capital: memory, chips, land, power, and shells must all be secured at the required pace.

Gerstner believes SpaceX has an unusual advantage in standing up physical infrastructure quickly. Friedberg similarly argues that Musk’s ability to build factories and industrial sites is a core competitive strength in an AI market that depends on data centers and semiconductor capacity. That execution advantage does not remove the macro risk. Gerstner calls the larger pattern “seller financing” or circular revenue: suppliers and financiers supporting buildouts whose end customers are a small number of frontier labs. If demand slips, compute-rental companies and semiconductor-linked equities can trade down together.

The potential answer to the financing question is Starlink. David Friedberg makes the strongest SpaceX bull case through Connectivity rather than compute rentals. Starlink generated $4.3 billion in quarterly revenue and $2.6 billion in adjusted EBITDA, according to the segment figures cited in the discussion. With 12 million subscribers—double the year-earlier figure—and a cited $66 in monthly average revenue per user, Friedberg described Starlink as a rapidly scaling cash engine.

Friedberg’s projection is that continued consumer and enterprise growth could bring Starlink toward $40 billion in annual revenue and roughly $30 billion in free cash flow. He argued that, at a 30-times multiple, Starlink alone could support a trillion-dollar valuation within roughly two years. Those are Friedberg’s bullish projections, not reported company guidance. Their function in the argument is to show how Connectivity could fund Starship, AI compute, and Terafab rather than compete with them for finite capital.

The Starlink expansion thesis also rests on Starship’s ability to increase network capacity. Sacks said Falcon 9 launches about 27 V2 satellites at a time, adding roughly 2.6 to 2.7 terabits per second of total capacity. Starship, he said, could deploy 60 V3 satellites in one launch, adding 60 terabits per second—more than 20 times the capacity added by a Falcon 9 launch. The displayed Starlink material described V3 satellites as providing ten times the bandwidth of V2 satellites.

That capacity matters because it is the path the discussion identified toward broader telecom ambitions, including direct-to-cell service. But it also reinforces the capital-allocation question. Starlink is the comparatively established cash-producing engine; Starship, Terafab, model development, and compute rentals are larger and riskier uses of the proceeds.

Airtable presents the inverse problem: a business with meaningful revenue but a capital structure built around expectations it did not meet. Bending Spoons agreed to acquire Airtable for $1.285 billion. Calacanis put the transaction value at roughly $2.25 billion including cash, against a 2021 peak valuation of $11.7 billion. Airtable had about $480 million in annual revenue, nearly $1 billion in cash, profitability, and approximately 20% annual growth.

Before the acquisition, Airtable spun out its AI-agent business, Hyper Agent. Sacks interprets the separation as a division between a venture bet and a private-equity-style operating asset. The founders and technical talent can pursue a new AI company, while Bending Spoons can manage Airtable as an established product.

The key operating signal for Sacks is the reported 30% sales-team quota attainment. His inference is that Airtable had a viable product-led-growth business, but that high-valuation investors pressed it toward a sales-led motion designed to produce venture-scale growth. The sales force did not generate enough incremental growth to justify its cost.

Sacks’s aggressive hypothetical is that Bending Spoons could strip out 80% to 90% of the cost base, retain much of Airtable’s product-led growth, and generate roughly $400 million in annual EBITDA—repaying the acquisition price within several years. That is Sacks’s scenario for the buyer, not an Airtable forecast.

Gerstner is more skeptical that the transition can be so clean. A slowing software company can suffer collapsing morale, rising customer churn, and staff departures to AI firms, he said. The product may also become less compelling if investment falls. In his view, a deal that appears cheap on revenue can look costly when measured against sustainable free cash flow; he suggested Airtable may have sold for closer to 30 times free cash flow.

The more important constraint is institutional. Founders and venture investors may be poorly suited to run a business in private-equity mode because high profitability can require dismantling the teams and ambitions they built. The cap table can compound the problem: liquidation preferences mean late-stage investors must be repaid before common shareholders participate in upside, which makes a long, modest cash-flow recovery less appealing.

Sacks argues that AI could make leaner maintenance more feasible. Models can reconstruct and work with a legacy codebase, reducing dependence on human institutional memory. That may give portfolio operators an advantage in maintaining mature products with much smaller teams.

Airtable is not, in Sacks’s view, evidence that all software faces the same fate. No-code tools are particularly exposed because their value involved replacing coding with another system users still had to learn. Claude Code, Lovable, and computer agents can reduce that learning curve by allowing users to describe what they want directly. Systems tied to identity, compliance, regulated work, and years of integrations are different. Microsoft and Salesforce are valuable, Sacks argues, because they are operational rails: access control, reporting, recorded communications, certifications, and embedded workflows are difficult to replace with a newly generated application.

Gerstner pointed to the IGV software ETF, which he said was up about 20% over six months and five years, and to Snowflake, which he said was up roughly 88% over six months. The implication is not that AI leaves software untouched. It is that the relevant divide is between products whose core function can be recreated cheaply and systems whose value lies in distribution, integration, compliance, and institutional trust.

The dispute over American training data turns on whether expert-generated data is strategic technology or a reproducible input.

Forbes reported that Silicon Valley data companies supplying OpenAI, Anthropic, and US government customers also sold training datasets to Chinese labs, including Tencent, ByteDance, Alibaba, and Moonshot. Calacanis said China’s top six AI labs were spending $500 million annually on what the report called “secret sauce”: PhD-authored content, reinforcement-learning pipelines, and expert knowledge.

Calacanis argues this is materially different from ordinary data labeling. In his description, specialists in coding, biology, science, and other technical fields evaluate difficult model responses, identify errors, write better answers, and verify the results. Selling the resulting packages to Chinese labs, he argues, gives those labs an avoidable shortcut in catching up with American systems.

He cited his use of Kimi, Qwen, and GLM 5.2 over the preceding 60 days as evidence that Chinese models have become highly capable. He believes expert-data sales, alongside distillation, are part of that convergence. He also said Micro One, a company in which he has invested, chose not to sell its data to China.

David Sacks does not reject export restrictions as a category. His test is whether a resource is genuinely proprietary, has dual-use or military relevance, and would materially affect the strategic balance. He pointed to restrictions on EUV lithography exports to China as an example of a targeted control that, in his view, had real strategic force.

Sacks doubts that expert data creation automatically clears that bar. China has substantial math and science talent, he says, and can create its own datasets. A ban on American suppliers could create reciprocal restrictions without preventing the underlying work. He favors reducing American dependencies on China, but argues that the countries still have meaningful interdependencies and that controls should be targeted at areas where they will have a real effect.

The disagreement is direct. Calacanis says recreating Western expert datasets at scale would require Chinese labs to recruit the best scientists and specialists from the West; in his view, the data packages transmit expertise that should remain an American advantage. Sacks replies that China’s scientific workforce is large enough that it is not obvious the country cannot reproduce the material internally. He is open to restrictions on genuine secret sauce but does not assume every proprietary dataset is decisive.

Brad Gerstner puts the dispute in a political context. He agrees with Sacks that the US benefits from competition and argues that America is currently ahead through frontier labs, domestic open-source development, and relatively limited regulation. That lead, he says, is why data sales alone are unlikely to produce a major policy shift.

But he expects continuing scrutiny in Washington. If officials conclude that China has caught up with or passed US frontier labs, transactions involving expert data, distillation, and chips will receive much more attention. The current tolerance, in Gerstner’s framing, rests less on a settled judgment that the data is harmless than on confidence that the US still leads.