How much AI token should a company allocate to its employees annually?

Jensen Huang's answer might shock many CFOs: $250,000.

By this standard, NVIDIA's annual token budget for its engineering team reaches as high as $2 billion.

On the flip side, Uber has taken a completely opposite approach.

In April this year, it burned through its entire AI coding budget for 2026 ahead of schedule. Subsequently, the company imposed a hard cap on engineers: $1,500 per person per month.

Forbes even crunched the numbers and came up with a particularly piercing headline:

"The cost of artificial intelligence has surpassed that of the talent it replaces."

This is intriguing.

Among the top tech companies, all purchasing AI, one dares to allocate $250,000 per person in budget, while another starts pinching pennies down to $1,500 per month. How did the gap widen to this extent?

Recently, overseas investment firm Leonis Capital wrote an article specifically discussing this phenomenon. Their answer is fascinating:

When companies buy AI, they are actually operating under two entirely different economic models.

One treats AI as a cost, aiming to accomplish the same tasks with less money. The other views AI as growth capital—as long as it creates more products, more customers, and more revenue, the more tokens spent, the better the deal.

Understanding these two economic models may be closer to the true core of AI commercialization than debating which model is stronger.

/ 01 / When Intelligence Becomes the Bottleneck of the Business Model

From Leonis Capital's perspective, two paradoxes lie at the heart of the issue:

First, why do some companies slash AI costs while others spare no expense on the most expensive cutting-edge models?

Second, if open-source models are closing in on the frontier, why can closed models still command such high premiums?

Behind these two questions lies the same fundamental issue—is intelligence the bottleneck of the business model?

Two answers correspond to two entirely different economic models.

Expansion Market: Output grows with intelligence. Higher intelligence directly translates to greater revenue. Buyers will pay top dollar for the best models available, regardless of price.

This is akin to Formula 1 racing. Spending $1 billion on cutting-edge AI is like buying the fastest car on the track, where lap time is the sole objective. Performance is everything—the better it performs, the more you're willing to pay.

Efficiency Market: Output is capped by non-intelligence constraints. Once intelligence surpasses the threshold set by other constraints, no amount of additional intelligence can generate extra output. In this case, buyers will only pay the price of the labor being replaced.

AI's role here is marginal: replacing labor, processing work orders, extracting fields, summarizing documents, handling claims. Once the model is competent enough, additional intelligence is meaningless.

The ultimate result is that expansion markets will move toward outcome-based pricing—revenue sharing in drug discovery, profit sharing in trading gains—resembling consulting or services rather than infrastructure.

Efficiency markets, meanwhile, will trend toward infrastructure logic, with pricing based on token or API calls, transparent caps, and revenue models akin to utilities like electricity and water.

Thus, among application layers serving both types of buyers, the same revenue should command vastly different valuation multiples.

/ 02 / Four Trends Behind the Token Frenzy

At this point, many might think: Why not just categorize tasks?

For example, "customer support = efficiency, coding = expansion." But this path is fraught with difficulties.

Because the same task can belong to different markets, depending on what the buyer wants to achieve with it.

Software engineering is the most typical example. Uber engineers generate about 70% of their submitted code via AI, yet the company is still desperately trying to limit spending—because passenger demand on ride-hailing platforms is unaffected by AI. More complex code doesn't translate to more orders.

But for an early-stage startup, ten times the code volume can indeed translate to ten times the products and ten times the customers.

Thus, an ironic situation emerges: high-growth startups are more willing to pay for cutting-edge coding than large enterprises with budgets a thousand times bigger.

The same company can even operate under both logics simultaneously.

Coinbase revealed its strategy in June this year: token usage is still rising, but total AI spending has been nearly halved—defaulting to cheaper open-source models and routing tasks to "the cheapest model that meets requirements."

But on its new prediction markets business line, the strategy is the opposite: betting that cutting-edge models can expand the market itself.

The same company, opposite goals, and diametrically opposed spending logics.

Under this framework, one conclusion is almost unavoidable: nearly all Fortune 500 companies are efficiency buyers.

Their businesses are constrained not by AI but by demand, distribution, regulation, physical assets, network effects, headcount, and all the other factors that truly drive large companies.

So after the "token maximization" frenzy, most companies will revert to efficiency-driven approaches. Currently, 98% of practitioners are managing AI spending, up from 63% in 2025; 73% of enterprises exceeded their AI cost expectations in the past year.

This means efficiency buyers once overpaid for efficiency gains and are now pulling back. From this, several judgments can be drawn:

First, the next phase of tokens will revolve around "allocation."

Companies will treat AI as an investment portfolio: model access will no longer be a universal right across the company but a tiered strategy. Performance metrics will shift from seat counts and benchmark scores to cost per task and value per dollar.

Second, open-source models have their moment—and it's bigger than expected.

Once task thresholds are crossed, GPT-5.6, Kimi 3, and GLM 5.2 deliver nearly identical commercial value.

Third, closed-source frontier labs must escape efficiency markets.

A significant portion of their revenue comes from tasks like coding, which are efficiency tasks for buyers—and this space is being eroded by open-source alternatives.

Anthropic's roadmap reflects this pressure directly: Claude for Financial Services in July 2025, Claude Science in June 2026.

From coding to finance to pharmaceuticals. This roadmap is all about escaping efficiency markets.

Fourth, efficiency markets are pricing like infrastructure.

Pricing will ultimately move toward token- or API call-based models, with caps and observable billing.

Competition will shift downward—inference costs, deployment, customization, workflow integration, and the mundane monitoring tools that let companies actually control AI spending. This is why OpenRouter achieved $50 million in annualized revenue and attracted billion-dollar acquisition offers.

It's essentially betting that efficiency markets will be the largest segment—and currently severely underserved.

/ 03 / Competition Dynamics Determine the Sustainability of Large Model Revenues

Because in expansion markets, the value of intelligence has no upper limit.

Bridgewater's AIA Macro Fund is an AI-run fund: AI agents trade, with human oversight. Starting with about $2 billion at the end of 2023, it grew to about $4.5 billion by mid-2026—an 11.3% annualized return, on par with its flagship human-run Pure Alpha fund.

When alpha can be directly attributed to intelligence, model quality is the only variable, and token price is irrelevant. But expansion buyers never purchase in isolation—they compete with each other.

And the nature of that competition determines the sustainability of labs' revenues.

One type is zero-sum. Greater intelligence merely redistributes a fixed pie.

Alpha in hedge funds is the classic example: one firm's gain means another's loss. So when one adopts a cutting-edge model, rivals have two choices—follow suit or drop out.

But when all funds use the same model, the advantage disappears. Total output remains unchanged, while industry-wide costs rise permanently.

Consulting is more subtle but in essence the same: if McKinsey adopts a model, BCG must follow, yet consulting budgets don't increase by a penny.

This spending is effectively a tax levied by competitors on each other. And model providers collect that tax.

A similar story played out in high-frequency trading.

In 2010, Spread Networks spent ~$300 million laying an 827-mile fiber cable through the Allegheny Mountains, reducing round-trip time between Chicago and New Jersey from 17ms to 13ms.

All funds had to follow suit or fall behind, so tower operators and exchanges became the biggest winners.

Later, microwave replaced fiber, and Spread Networks was acquired by Zayo for $131 million—less than half its construction cost.

This zero-sum spending has two characteristics:

It's extremely intense—as long as competition persists, no one dares stop. So despite complaints from every advertiser, ad tech spending has kept rising for two decades.

It's extremely fragile—it collapses instantly when structural conditions change. When exchanges erected barriers, the HFT arms race ended abruptly. Consolidation, regulation, or a tacit truce would have the same effect on any zero-sum AI race.

The second type is positive-sum. Greater intelligence expands the total pie.

Drug discovery is the clearest example. If every company doubles its success rate, more drugs reach the market. Competitors jointly create more cures and more total revenue.

GLP-1 drugs generated ~$132 billion in sales in 2025, up over 30% YoY—a pie so large that the compute costs of discovering the next molecule are negligible.

Chip design and materials science follow the same logic. A better codebase doesn't degrade rivals' codebases. This looks more like investment than taxation.

Positive-sum spending can't be easily disrupted by external forces, but it will eventually hit non-intelligence bottlenecks—drug discovery spending will keep rising until it hits clinical trial capacity or regulatory approval ceilings. Then, no matter how strong the model, spending will plateau.

For frontier labs, here's a strategic conclusion:

Composition matters more than total.

Two equally profitable labs—one reliant on zero-sum clients, the other on positive-sum clients—are fundamentally different businesses. The former is like tax collection: stable but fragile. The latter is like investment: durable but hits walls.

No one has quantified this mix yet. It's the most overlooked metric in AI—we suspect labs haven't calculated it either.

/ 04 / Conclusion

When you piece the two paradoxes back together, the entire discussion around AI spending takes on a completely different shape.

The market keeps asking: Is AI worth what companies are paying?

That question is fundamentally flawed. It averages two markets that should never be averaged in the first place.

The only correct question is: For every dollar of AI investment, is intelligence the bottleneck for what it's trying to produce?

If yes—frontier AI is cheap at almost any price.

If no—the cheapest compliant model is the only rational choice, and buyers paying a premium are just overspending.

But this conclusion makes almost everyone uncomfortable.

Frontier labs are uneasy: a significant portion of their revenue hangs by a thread, depending on whether they can outrun open-source models into new expansion markets.

Open-source labs are uneasy too: they're targeting the largest yet least profitable segment, with success hinging on whether they can capture efficiency markets at commodity-level margins.

Companies are uneasy: most are efficiency buyers, and any AI project not tied to real bottlenecks is just performance art masquerading as strategy.

The AI narrative itself is uneasy: the thrilling story of "AI transforming every function of every company" only holds true in a minority of cases.

So for the tiny fraction of markets truly constrained by intelligence, the really interesting questions have little to do with cost:

What happens when "usable intelligence levels" finally cease to be the limiting factor for human output?

Most of the AI market still aims to reduce costs, and that's fine. But what truly matters is where the value of higher intelligence keeps climbing.

That's the frontier. That's why AI exists.

By Qi