Prediction Markets Are Building a Yield Curve for AI Compute

The most interesting AI market right now may not be Nvidia stock, OpenAI's next model, or another debate about whether inference costs are falling.

It may be this: prediction markets are now putting a probability distribution on the end-of-month rental price of a B200 GPU.

That sounds narrow. It is not. B200 rental prices sit close to the center of the AI infrastructure trade: GPU supply, data center power, hyperscaler capex, neocloud utilization, training budgets, inference margins, and the question underneath the whole cycle: is frontier compute still scarce, or is it becoming a commodity?

As of June 10, 2026 at 6:54 p.m. PT, SimpleFunctions market data showed Kalshi pricing a 93% chance that B200 compute per hour finishes above $4.07 by June 30, but only a 14% chance that it finishes above $6.27. Polymarket's B200 end-of-June bucket market put the highest displayed price on the $5-$6 range, with smaller but meaningful prices on $4-$5 and $6-$7.

That is not a "compute is crashing" curve. It is also not a "GPU scarcity goes vertical forever" curve.

It is a tradable band.

Disclosure: this is market analysis, not investment advice; SimpleFunctions holds no position in these contracts.

The Curve

The cleanest way to read the market is not from any single contract. The single-contract prices are noisy, liquidity is thin in places, and some strike prices have wide spreads. The better read is the shape of the distribution.

On Kalshi's June 30 B200 ladder, SimpleFunctions scan data showed:

The non-monotonic points matter. A clean options surface should not have odd kinks like that. But thin prediction markets often do, especially when different strikes have different order books, attention, and stale limit orders. The correct response is not to pretend the curve is perfectly efficient. The correct response is to extract the broad signal and respect the microstructure limits.

The broad signal is clear enough: the market strongly rejects a sub-$4 B200 print, sees the mid-$4 to mid-$5 region as live, and treats the high-$6 to $7-plus zone as a tail.

Polymarket tells a similar story in bucket form. Its B200 end-of-June market, which resolves on the finalized Ornn B200 Index value for June 30, showed these raw displayed outcome prices in the SimpleFunctions scan:

Do not normalize that table too literally. The outcome prices are not a clean institutional probability surface: several buckets have wide spreads, and raw displayed prices can carry stale prints or order-book noise. The useful point is simpler. The modal outcome is $5-$6, the adjacent $4-$5 bucket is alive, and the very low and very high tails are much weaker.

This is the point: we now have a market-implied AI compute curve.

Why B200 Pricing Matters

B200 is not just another GPU ticker. Nvidia's Blackwell platform is the current high-end symbol of the AI infrastructure race. Nvidia describes the GB200 NVL72 system as a rack-scale, liquid-cooled design connecting 36 Grace CPUs and 72 Blackwell GPUs for large-scale LLM inference, and says the system is built for trillion-parameter inference workloads. That puts Blackwell pricing close to the marginal cost curve for frontier AI capacity.

A falling B200 rental price would mean something. It could signal supply finally catching demand, neocloud competition compressing margins, or customers getting better access to top-end accelerators without long-term commitments.

A rising B200 rental price would also mean something. It could signal that demand is still outrunning delivered supply, that power and data center constraints are binding, or that buyers continue to pay up for the latest generation of compute rather than shift down to older chips.

This is why the B200 curve is more interesting than most AI headlines. It compresses a lot of vague debate into one observable question: what is the market clearing price for the highest-demand compute?

The Spot Reference

The prediction markets are not floating in a vacuum. Ornn's public API documentation lists B200, H200, H100 SXM, A100 SXM4, and RTX 5090 as supported GPU index series. Ornn describes itself as building the market layer around compute: benchmarks, market data, and risk-transfer products.

The Ornn B200 history is volatile. Its public API showed:

That path is not a simple deflation story. B200 rose hard from March into late May, then pulled back into June. As of the latest data point, it was still materially above the March starting level.

Public cloud pricing gives another useful anchor. Runpod's pricing page listed B200 SXM at $5.89/hr, H200 at $4.39/hr, and RTX 5090 at $0.99/hr. Those are not perfect substitutes for an index. They reflect one provider's listed on-demand pricing, not a full market-clearing composite. But they line up with the market's broad B200 distribution: not below $4, not obviously above $7, and very plausibly in the $5-$6 zone.

This is where the market signal becomes useful. The prediction market is not merely guessing a number. It is aggregating views about where a benchmarked compute index will land relative to visible spot prices, recent index history, and trader expectations about supply and demand into month-end.

The Wrong Take Is "Compute Is Crashing"

There is a tempting LinkedIn version of this story: AI compute is getting cheap, GPUs are commoditizing, and the infrastructure bubble is over.

The data does not support that.

Business Insider reported in April that Silicon Data's Neo Cloud B200 index climbed from $4.40 to $5.35 over three months, a 22% increase. The same report framed the move as evidence that AI compute demand was still outpacing supply, with constraints not just in chips but also memory, power, and data center capacity.

The Ornn B200 series since March also does not look like a collapse. It looks like a volatile, supply-constrained commodity beginning to trade in a range.

That distinction matters. A commodity can become more transparent before it becomes cheap. Oil has a forward curve. Power has a forward curve. Memory has cycles. Freight has rates. Compute is starting to look similar.

The existence of a curve does not mean the asset is no longer scarce. It means the scarcity is becoming legible.

The Macro Backdrop

The reason this curve matters is that the AI capex machine is enormous.

Axios reported that Alphabet, Amazon, Meta, Microsoft, and Oracle had raised $255.34 billion through debt and equity by early June 2026, more than twice the full-year 2025 amount, and that the five companies have said they may spend roughly three-quarters of a trillion dollars on AI data centers by year-end.

Business Insider separately reported Bank of America's framing that investors should own AI "capex takers" such as semiconductors and hardware, while becoming more cautious on the largest AI spenders as capex draws down free cash flow. The same report said the five large hyperscalers were expected to spend $725 billion this year.

Whether those numbers prove brilliant or excessive depends partly on utilization and pricing. If compute stays scarce and high-priced, the owners of capacity can justify enormous buildouts. If rental prices fall faster than expected, the capex spenders face a harder margin story. If prices remain volatile, the ability to hedge compute exposure becomes valuable.

That is why a B200 prediction market is not a novelty. It is a small public window into the marginal economics of the AI buildout.

What The Market Is Really Saying

The current curve implies four things.

First, traders do not expect a near-term collapse in B200 rental prices. Kalshi's 93% price on the $4.07 threshold and Polymarket's low displayed price on the under-$4 bucket both point in the same direction.

Second, traders also do not appear to believe the late-May spike is guaranteed to persist. Ornn's B200 index touched $6.11 on May 30, but Kalshi priced the $6.27 threshold at only 14% and Polymarket put the $6-$7 bucket far below the $5-$6 bucket.

Third, the market sees the center of gravity around $5. That is consistent with Ornn's latest B200 value of $4.79 and Runpod's public B200 price of $5.89/hr.

Fourth, this market is still young. Spreads are wide in several Polymarket buckets and some Kalshi strikes are visibly noisy. This is not yet a clean institutional futures curve. It is a rough early curve built out of prediction market contracts.

But early curves are often messy. That does not make them useless.

Why This Belongs On A Prediction Market Platform

Prediction markets are usually discussed through elections, sports-adjacent events, macro releases, or crypto outcomes. AI compute is different. It is closer to a physical commodity market hiding inside the software economy.

The market structure is natural:

• There is a measurable reference index.
• There is genuine uncertainty.
• There are participants with different information sets.
• There are users with real exposure.
• There is a broader audience that cares about the result.

AI labs care because rental costs affect training and inference budgets. Neoclouds care because utilization and pricing determine payback periods. Investors care because GPU economics flow into capex returns. Model companies care because input costs shape gross margins. Infrastructure buyers care because timing purchases incorrectly can be expensive.

A prediction market can turn those views into prices.

The interesting part is not whether one June contract resolves at $4.91 or $5.08. The interesting part is that the market is beginning to price compute as something that can be watched, traded, and eventually hedged.

What Would Change The Read

The bullish scarcity case would strengthen if B200 index values rebound back above $6 and the upper Kalshi strikes tighten with real bid depth. That would suggest the early-June pullback was temporary and that the market had underpriced renewed supply pressure.

The deflation case would strengthen if the Ornn index breaks below $4 and stays there. That would make the current market distribution look too conservative and would support the view that new B200 supply or competitive neocloud pricing is finally hitting the spot market.

The microstructure warning would strengthen if the strike ladder remains non-monotonic and spreads stay wide. In that case, the curve is still more of a directional sentiment surface than a precise market-implied distribution.

There is also a reference-risk point. These markets resolve on index methodology, daily finalization, and venue-specific rules. Anyone interpreting the prices should understand the contract terms, not just the headline GPU name.

Our Read

The market is telling a more subtle story than the public AI debate.

B200 compute is expensive. It is no longer unknowable. And it is not yet obviously collapsing.

The best read from current prices is that traders expect the June 30 B200 index to land in a bounded premium range, with the center of mass around $5-$6/hr and the tails fading below $4 and above $7. That makes sense against Ornn's latest $4.79 print, Runpod's $5.89 public B200 price, and the broader backdrop of heavy AI infrastructure spending.

The bigger conclusion is structural: AI compute is becoming financialized.

That does not mean every GPU will trade like crude oil. It does mean that the market is starting to build the primitives: reference indexes, dated contracts, bucket distributions, strike ladders, and visible probabilities.

For an industry spending hundreds of billions of dollars to turn electricity, chips, and data centers into intelligence, that is a real development.

The AI market has spent years arguing about model capability. The next layer is cost.

And now the cost has a curve.

Live markets: Kalshi B200 contracts for above $4.07, above $4.87, above $6.27, and above $6.87; Polymarket's B200 end-of-June bucket market. Prices move; the numbers above are timestamped June 10, 2026 at 6:54 p.m. PT.

FAQ

What are AI compute price odds?
They are prediction market prices that imply probabilities for future GPU rental prices, such as whether the B200 compute price will finish above a specific dollar-per-hour threshold.

What is the market implying for B200 rental prices at the end of June 2026?
As of June 10, 2026 at 6:54 p.m. PT, the market-implied center of gravity was around the $5-$6 per hour range, while the tails below $4 and above roughly $7 were priced much lower.

Why do B200 GPU rental markets matter?
B200 rental prices are a live signal for high-end AI compute scarcity, neocloud pricing power, hyperscaler capex returns, and the marginal cost of training and inference capacity.