AI Data Bottleneck: The Expensive Constraint Behind the Capex Boom

This is not a small-company story. As of June 2026, CompaniesMarketCap put Nvidia (NVDA) at about $5.05 trillion, Alphabet at about $4.26 trillion, Microsoft at about $2.73 trillion, Amazon at about $2.50 trillion, Tesla at about $1.52 trillion, and Meta at about $1.43 trillion. That is the market’s AI bill, priced through the biggest public companies in the trade.

The problem is not that AI does not work. It works.

The problem is that the market is treating scale as if it solves everything. Buy more chips. Build more data centers. Train larger models. Sell more tokens. That is the clean version of the story.

The dirty version is data.

Not generic internet text. Not another scrape. The constraint is verified, domain-specific examples, expert labels, reward models, rubrics, and feedback loops. Compute can be bought. Power can be contracted. Data like that has to be made.

That is slower. It is messier. And it decides who gets paid.

Key facts at a glance

• Epoch AI measured the best open-weight model at roughly a 4-month lag behind the best closed model in its May 28, 2026 update, with an ECI gap near 8 points, according to Epoch AI.
• The original Chinchilla paper argued that compute-optimal training needs roughly 20 tokens per parameter, according to Hoffmann et al., published March 29, 2022.
• Human sample efficiency estimates are ugly for AI. One analysis puts the human advantage at roughly 100 to 10,000 times on some task comparisons, and as high as 1 million times by token-count comparison, according to Dwarkesh Patel, published in 2024.
• Data labeling is already a multibillion-dollar business. Grand View Research estimated the global data collection and labeling market at $2.38 billion in 2024 and projected $17.10 billion by 2030, according to Grand View Research.
• Precedence Research estimated the AI data labeling market at $3.77 billion in 2025 and projected $18.23 billion by 2035, according to Precedence Research.
• Public options data as of June 18, 2026 showed NVDA call volume of 2.39 million contracts versus 1.29 million put contracts, a put-call volume ratio of 0.54. The options market was not acting scared.
• Polymarket showed Anthropic at 84% odds to have the best AI model at the end of July 2026, versus 11% for Google and 5% for OpenAI, according to Polymarket checked on June 22, 2026. That is a betting market, not proof. But it shows how narrow the perceived frontier has become.

The market is buying scale

Start with the tape.

The AI complex is not priced like a science project. It is priced like infrastructure, software, advertising, cloud, autonomy, and semiconductors all found a new growth engine at the same time.

NVDA is the seller. MSFT, GOOGL, AMZN, and META are the buyers. TSLA is the data-scale outlier because driving data is one of the few places where a public company can point to a giant proprietary behavioral dataset.

Public options data as of June 18, 2026 backed up the lack of fear. NVDA had 2.39 million call contracts trade versus 1.29 million puts, a 0.54 put-call volume ratio, with at-the-money implied volatility near 35.7%. AMZN had 816,987 calls versus 316,938 puts, a 0.39 ratio, with at-the-money implied volatility near 31.3%. META had 390,082 calls versus 225,962 puts, a 0.58 ratio, with at-the-money implied volatility near 33.7%.

Calls beat puts across the main AI buyers and sellers. The market is not hedging the story. It is still paying for it.

That is exactly why the data bottleneck matters.

The bottleneck is not a metaphor

The old AI debate was chips. The current AI debate is power. Both matter. But neither is the deepest constraint.

A model does not become a medical expert because it saw more web pages. It needs clinicians judging examples. A legal model needs lawyers. A coding model needs engineers. A research model needs evaluators who know when an answer is not just fluent, but right.

The Chinchilla result is the simplest way to see the issue. More parameters help. More data helps. But they are not the same thing. The Chinchilla paper said compute-optimal training required roughly 20 tokens per parameter. That finding made the field more data-hungry, not less.

And the human comparison is brutal. Humans do not need trillions of tokens to become useful. AI systems often do. The sample-efficiency analysis puts the gap at roughly 100 to 10,000 times on some task comparisons and up to 1 million times by token-count comparison.

You do not have to accept the top of that range to see the point. Take the low end. It is still bad.

The bull case says the spend becomes a moat. Maybe. But that only works if the data is scarce, the labeling pipelines stay human-heavy, and the open models do not close the gap fast enough.

Those are conditions. Not facts.

The open-model lag cuts both ways

Epoch AI’s May 2026 update is the cleanest public number here. The best open-weight model lagged the best closed model by roughly 4 months, with an ECI gap near 8 points, according to Epoch AI.

That number can be read two ways.

The optimistic read for the big labs is simple. Open models are close, but not equal. The last stretch matters. If the final 5% of capability needs expensive private data, proprietary feedback loops, and expert review, the top labs can keep the lead.

The skeptical read is just as important. A 4-month lead is not a fortress. It is a quarter. If open models keep sitting a few months behind the frontier, the model layer starts to look less like a monopoly and more like a fast-decaying product cycle.

That is the contradiction. The data bottleneck can strengthen the winners. It can also expose how short the product lead really is.

The table that matters

This table separates the facts from the interpretation. That distinction matters because the market often treats them as the same thing.

Evidence	What the data says	What it proves	What it does not prove
Open-model lag	Best open-weight model roughly 4 months behind best closed model in May 2026, per Epoch AI	Frontier labs still have a measurable lead	The lead is durable
Scaling law	Roughly 20 tokens per parameter in compute-optimal training, per Hoffmann et al.	Data volume matters directly	Bigger models alone solve sample efficiency
Sample efficiency	Humans estimated roughly 100 to 10,000 times more sample efficient in some task comparisons, per Dwarkesh Patel	AI still learns expensively	The gap is permanent
Labeling market	$2.38 billion in 2024 to a projected $17.10 billion by 2030, per Grand View Research	Expert data is becoming a real cost center	Revenue for labelers equals moat for model labs
Options tape	NVDA put-call volume ratio of 0.54 on June 18, 2026, based on public options data	Traders were not paying heavily for downside protection	Options positioning validates the AI thesis

The table says the constraint is real. It does not say the constraint automatically becomes a moat.

That is the part investors skip.

Expert data is the cost center

The data-labeling market numbers are not huge compared with AI capex. That is the point.

If a $2.38 billion market in 2024 grows to $17.10 billion by 2030, as Grand View Research projected, the spend is still small relative to the hundreds of billions going into chips and data centers. But it is the narrow spend. It is the spend that decides whether the hardware has anything useful to learn from.

Precedence Research put the AI data labeling market at $3.77 billion in 2025 and projected $18.23 billion by 2035, according to Precedence Research. Different base. Same direction.

The market is saying this is an infrastructure race. The evidence says it is also a judgment race. Who can source the best experts? Who can turn those experts into rubrics? Who can verify outputs without fooling themselves? Who can repeat the process in medicine, law, coding, finance, science, and operations?

That is not a chip order. That is an operating system.

Reinforcement learning is the escape hatch

There is a real counterargument.

Human labels are expensive. AI feedback is cheap. Reinforcement learning from AI feedback may replace part of the human pipeline. The Labelbox comparison describes the promise clearly: use AI systems to evaluate outputs that humans used to label.

If that scales, the data bottleneck weakens. If a model can generate candidate answers, score them with a verifier, and train on the best rollouts, the system starts manufacturing its own training signal.

That is not magic. It still needs verifiers. It works best where answers can be checked, like code, math, and formal reasoning. It is weaker where the target is judgment, taste, strategy, medicine, law, or messy human context.

But it is enough to make the moat argument fragile.

The bulls say data scarcity protects the incumbents. Maybe. But if AI feedback replaces enough human feedback, the scarcity gets diluted. The moat becomes a cost curve. Cost curves can move.

What this does not tell you

This does not tell you that AI is a bubble. That would be too easy.

It also does not tell you that NVDA, MSFT, GOOGL, AMZN, META, or TSLA are bad businesses. Public prices on June 22, 2026 show the market still assigning real value to the AI complex, according to Yahoo Finance chart data. Public options data as of June 18, 2026 showed call-heavy positioning in several of the same names.

This does not prove the top labs win. A 4-month open-model lag can be read as a lead. It can also be read as weak durability.

This does not prove expert labeling grows forever. A market projected from $2.38 billion in 2024 to $17.10 billion by 2030 is a serious growth market, according to Grand View Research. It is not proof that human labeling remains the binding constraint in 2030.

And this does not prove return on capex. That is the missing evidence. Revenue growth is visible. Infrastructure spending is visible. The question is whether the data and verifier layer turns that spending into durable margins.

The verdict

The AI data bottleneck is real enough to change the investment question.

The question is no longer whether models improve. They do. The question is whether improvement requires a scarce input that only the biggest labs can afford, or whether open models and AI feedback make that input cheaper every quarter.

If the scarce-input view is right, returns concentrate. The winners are the labs and platforms that can buy chips, hire experts, build verifiers, and absorb failed training runs.

If the cheap-feedback view is right, the model layer commoditizes faster than the spending cycle pays back. Then the AI capex boom still builds useful infrastructure, but the equity returns spread poorly.

That is the fact the market has not settled.

The market is paying for scale. The data says scale alone is not the question anymore. Returns on the spend are.