Two words sit at the center of the most forwarded research note of the week, and nobody — not the bank that wrote it, not the crypto outlet that recycled it, not the traders now arguing about it in public — has bothered to define them: token prices.
Ledger update: Capital is fleeing. The only open question is which ledger it is fleeing from.
Read that sentence one way, and it describes the collapse of per-million-token inference pricing across the large language model industry — the unit cost of machine cognition falling fast enough that the return on AI data-center capital expenditure is being compressed toward the floor. Read it another way, and it describes a basket of crypto AI tokens — TAO, FET, RNDR, IO, AKT — bleeding value because the narrative premium that funded their GPU supply networks is evaporating in a risk-off tape.
These are not the same story. They do not share the same balance sheet, the same payer, or the same timeline. One is a margin problem facing Nvidia's largest customers. The other is a business-continuity problem facing a handful of protocols whose operating capital is denominated in their own token. And here is what turns a semantics quibble into a tradeable signal: the headline was published by a crypto outlet, sourced to Goldman Sachs, tagged to "AI infrastructure," and left deliberately open. That is not sloppy editing. That is a keyword being asked to do two jobs at once, and the failure mode is that readers will position themselves against the wrong one.
I have spent the better part of a decade auditing exactly this kind of gap — the distance between what a headline promises and what the underlying data will support. In 2017, at the height of the ICO mania, I built a script to reconcile the EOS pre-sale's stated supply against live chain data and surfaced a 40 percent discrepancy that took the token down 15 percent in six hours, before the project paused the sale. The lesson was never that the numbers were wrong. The lesson was that the most dangerous part of any narrative is the vocabulary nobody stops to define. This Goldman note is that lesson, republished for 2026, and the market is walking straight into it.
So let me do the forensic work the headline refused to do. Both readings. Both balance sheets. Both cost structures. And the point where they touch — because they do touch, and the intersection is the single most underpriced signal in the market right now. Alpha dropped: Follow the money.
Context: How "AI Infrastructure" Became One Undifferentiated Trade
To understand why the word "token" could plausibly mean two things inside a single headline, you have to understand how two entirely separate capital cycles got welded together under one label.
The first cycle is real and enormous. Through 2024 and into 2026, the four largest US cloud operators — Microsoft, Google, Amazon, Meta — committed well over half a trillion dollars to AI data-center construction, GPU procurement, power contracts, and network interconnect. This is not speculation. It is steel, silicon, and megawatts, committed against multi-year depreciation schedules. The revenue side is where the model gets fragile. Inference — actually serving models to paying users — has been commoditizing faster than any of those capex plans assumed. When I built predictive liquidity models during DeFi Summer in 2020 and watched emission schedules eat themselves in real time, the pattern I learned was this: supply that arrives on a schedule always outruns demand that arrives on a decision. AI data centers are supply on a schedule — eighteen to thirty-six months from board approval to first token served. Token consumption is demand on a decision — a CFO renewing a contract, an enterprise flipping a workload on. The two clocks do not sync, and when they diverge, the spread is where the losses accumulate.
The second cycle is the crypto AI narrative. It began as a genuine response to a genuine bottleneck — GPU scarcity and hyperscaler pricing power — and it attracted real money: Render's decentralized rendering market, Akash's compute exchange, Bittensor's subnet incentive model, io.net's aggregated GPU pools. The pitch was clean and it landed: the world needs compute, the world's compute is centralized and expensive, therefore a token-coordinated supply network can undercut the cloud. Between 2023 and 2025 that pitch raised billions. By 2026, in a bear tape, it is being margin-called.
The analytical error is that these two cycles now trade under a single label, "AI infrastructure," and share a single narrative engine, "AI demand is exponential." That coupling is where the thesis breaks. AI demand being exponential does not mean AI infrastructure investment is profitable. It does not mean AI data centers earn their cost of capital. And it certainly does not mean AI tokens — of either variety — are worth what they were in 2024. The label is doing the work that the fundamentals should be doing, and in a bear market, labels get repriced first and hardest.
Core: The Two Ledgers, Separated
Let me pull the two readings apart cleanly, because the entire investment implication rides on which ledger you are actually reading from.
Reading One — The AI Inference Token: A Margin Collapse in Slow Motion
If "token prices" means the per-million-token price charged for LLM inference, then the Goldman warning is not a crypto story at all. It is a story about the unit economics of machine cognition, and it is ruinous for anyone who financed a data center against 2024 pricing assumptions.
The cost curve for running inference has fallen off a cliff, and it has done so through at least five independent mechanisms that compound on top of one another.
Model quantization and distillation shrink the compute required per token without proportionally degrading output quality. Mixture-of-Experts architectures activate only a fraction of total parameters per token, cutting the effective compute footprint per request. Speculative decoding lets a small draft model propose and a large model verify, raising throughput per GPU-second without changing output. KV-cache optimization and paged attention reduce memory overhead, which raises the number of concurrent sequences a single accelerator can serve. And custom silicon — Groq's LPUs, Google's TPUs, and a widening array of inference-optimized ASICs — attacks cost-per-token from the hardware layer, where the leverage is largest.
Stack those together and the observed decline in inference cost is not linear. It is closer to a halving every twelve to eighteen months on many workloads, and faster on the simplest ones. This is the fastest deflation in the history of commercial computing, and it is happening inside a cost structure that cannot deflate at all.
Here is the trap, and it is the same trap I flagged in DeFi incentive design six years ago: when your revenue is priced per unit and your cost is fixed, a falling unit price does not reduce your profit — it erases it. A data center financed in 2024 against revenue of $X per million tokens is a data center financed against a revenue line that now clears at a fraction of $X. The GPU does not depreciate any more slowly because the product got cheaper. The power purchase agreement does not renegotiate itself. The debt service does not care that the market repriced the output. The capex was underwritten at one price and is being serviced at another, and the gap between them is the ROI Goldman is quietly warning about.
This is where the Jevons paradox enters, and where most AI bulls make their stand. The argument is elegant: cheaper tokens drive more consumption, and total revenue grows even as unit price falls. In aggregate, this has been true — total token consumption has exploded. But Jevons demand is a demand-side comfort and a supply-side illusion. The increased consumption accrues to whoever serves at the lowest marginal cost. If your rival is an open-weight model running on commodity hardware, or a hyperscaler amortizing its capex across a dozen other businesses, you are not the beneficiary of Jevons. You are the casualty of it. Volume growth does not rescue a business whose per-unit margin has gone negative — it accelerates the bleed, because every incremental request is served at a loss.
The ROI arithmetic is unforgiving. If you spent one dollar to build capacity that now yields forty cents of gross profit per year, no amount of demand growth closes the gap unless that growth is captured at a spread which no longer exists. The thesis that "AI infrastructure is a growth market" is not wrong. It is incomplete. Growth in use is not the same thing as growth in return, and the market has spent two years pricing the former and ignoring the latter.
Reading Two — The Crypto AI Token: Operating Capital That Melts
The second reading is shorter, entirely mechanical, and almost certainly the one the crypto outlet wanted you to take.
These networks carry a structural feature that no white paper advertises and every supply-side participant understands: the token price is the operating capital of the network. A decentralized GPU marketplace does not pay its providers in dollars. It pays them in emissions, which the provider sells into the market to cover electricity, depreciation, and its own cost of capital. The token is not a governance ornament or a speculative sideshow bolted onto a working business. It is the payroll. It is the thing that turns a GPU owner's electricity bill into a reason to plug in.
Which means the causal chain runs in exactly one direction, and it is not the direction the narrative assumes:
When token prices fall, provider compensation measured in real terms falls with them. Providers, who are economically rational and structurally disloyal to any logo, disconnect. GPU supply contracts. Network capacity drops. Latency rises. The product degrades for whatever paying customers exist — and that is the part the pitch never confronts — because the token was subsidizing the supply side, not the demand side. The network was never paying customers to show up. It was paying suppliers to pretend the demand was there.
This is not a price decline. This is a business-continuity event wearing the costume of a price decline. It is the identical structure I dissected in 2020, when I built the model showing that 60 percent of high-yield DeFi protocols would face insolvency within three months of their emission schedules colliding with real demand. The yield was never real. It was inflation pointed at the supply side, and the moment the inflation stopped, the supply left. Decentralized compute is running the same playbook, two asset classes later, and the outcome mechanics do not change because the word "AI" is attached to them.
The value-capture fracture is total, and it is the part that should end the debate. Crypto AI tokens do not capture the value of the compute they coordinate, because compute is priced in dollars by a global market and the token does not sit in the payment path. A customer buying GPU hours from a decentralized network pays in stablecoins or fiat, not in the network token. The token's only function is to subsidize supply. It is a subsidy voucher dressed as an asset, and the moment the voucher loses value, the subsidy — and the network it funds — evaporates together. I have looked at the tokenomics of a dozen AI-crypto hybrids ahead of the 2026 convergence, and roughly 80 percent lack any utility beyond speculation once you strip the incentive layer away. That was my finding before this headline existed. The Goldman note does not reveal it. It merely makes it impossible to ignore.
The Intersection: Where the Two Ledgers Actually Touch
Here is why any of this matters beyond two competing hot takes. The two readings are not disjoint. They meet at a specific, listed group of companies that has spent the last two years quietly becoming the strangest hybrid in public markets: the bitcoin miners who pivoted into AI and high-performance computing data centers.
IREN, Core Scientific, Hut 8, TeraWulf, Cipher — firms that built their balance sheets on proof-of-work hashrate and their valuation narratives on AI hosting contracts. Their equity is priced against two variables at once: the price of bitcoin and the durability of AI compute-leasing revenue. If Reading One is correct — if inference economics are collapsing — then the hosting contracts those miners signed at 2024 rates are underwater before the accelerators are racked, and the AI premium embedded in their multiples unwinds. If Reading Two is correct — if the crypto AI narrative is deflating — then the capital markets that funded their pivots tighten, and the pivot stalls before it generates a single dollar of contracted revenue.
Either way, this group takes the hit, and it takes it twice, because it is the only place in the market where the two ledgers are wired into the same share price. That is the crossover the headline concealed, and it is the reason a Goldman view on "AI infrastructure ROI" is a crypto story whether the original author intended it to be one or not. Most crypto desks will not connect these dots. The connection is the trade.
What Goldman Actually Said, and Why the Chain of Custody Matters
I want to be precise about what is verifiable here, because forensic discipline is not optional when a note can move a sector.
What is confirmable: a Goldman-sourced view suggesting token prices are pressuring AI infrastructure returns, with open-source competition cited as a background variable, republished by a crypto outlet. Of the handful of extractable facts, four trace to a single paragraph. There is no named analyst. There is no publication date. There is no primary document. There is no framework under which the word "token" is defined.
What is not confirmable: whether this is new information, which token market it references, whether it originates in a public report or a client-only note, and whether it was authorized for distribution at all. A secondhand summary of an undated, unattributed research view is not analysis. It is a lead. And I have learned — from verifying EOS supply claims on-chain in 2017 to tracing NFT wash-trading wallet clusters in 2021, where I mapped the address groups controlling 70 percent of a collection's volume — that the credibility of a market-moving claim is only ever as strong as the weakest sentence in its chain of custody. Here, the weakest sentence is the headline itself.
This is not a reason to dismiss the view. Goldman's research desk carries real weight, and the underlying thesis — that AI capex ROI is under pressure — is defensible on independent technical and economic evidence, as I have laid out above. It is a reason to trade the analysis, not the headline. The headline is ambiguous by design. The analysis, once you reconstruct it, is unambiguous: pricing power is migrating away from the entities that spent the capital to build the capacity. That migration is the entire story, and it does not care which definition of "token" you bring to it.
Contrarian: Free Is the Enemy, and Everybody Is Pricing Demand
Here is the angle almost nobody is trading, and it is the one I would build a position around.
The consensus has spent two years pricing AI as a demand story. More users, more queries, more agents, more tokens. The demand data supports it completely. Total inference volume is compounding. Model capability is improving. Enterprise adoption is real. Every single input to the AI demand thesis checks out, and that is precisely why the consensus is crowded on the wrong side of the return.
The consensus has spent almost no time pricing the supply side's inability to hold a price. That is where the return lives or dies. The single most destructive sentence in the AI economic conversation is this: an open-weight model that is 90 percent as good as the frontier model, at zero marginal license cost, permanently caps the price of the last 10 percent. Llama, DeepSeek, Qwen, Mistral. These are not competitors in the traditional sense. They are a price ceiling enforced by physics and by the fact that their weights are public. You cannot out-earn a rival whose product costs you nothing to distribute and whose marginal price is a legal zero.
Open-source will never stop lagging permanently, and the bulls will keep saying it will. It doesn't. The gap between open-weight and frontier models has been compressing on every benchmark that matters — coding, reasoning, long-context retrieval — and each compression event re-prices the entire closed-model inference stack downward. When an open-weight model enters the top ranks of the public leaderboards outright, it does not merely take market share. It deletes the pricing power of every model above it, and it deletes the ROI of every data center financed beneath it.
Apply that lens to the crypto AI tokens and the question gets brutal fast. If a competitive open-weight model runs on commodity hardware, if inference is racing toward zero, and if the decentralized network's only differentiator is a token subsidy, then what is the token for? Three years of "AI plus crypto" decks have never answered this cleanly. It is the same unanswered question that has kept soulbound tokens a concept for three years — nobody wants their identity permanently on-chain, and nobody wants to pay a subsidy premium for compute they can rent cheaper from a hyperscaler with a real SLA and a real support contract. The moat was never the technology. It was the novelty of the pitch, and novelty depreciates faster than any GPU.
The contrarian trade is not "short AI." It is long the demand while short the pricing power. Demand is real and durable. The returns accrue to whoever controls the scarcest input — and in a world where model weights are free and inference is commoditized, the scarcest input is not the model and it is not the token. It is power, land, interconnect rights, and a signed lease with a creditworthy counterparty. Everything else is a margin waiting to compress, and the market has not yet repriced that distinction.
Risk Assessment: What Confirms This, and What Breaks It
A thesis without falsification conditions is a vibe. Here are the numbers that will settle the argument, both of them public.
The first is hyperscaler capital expenditure guidance. When Microsoft, Google, Amazon, or Meta next revise their AI capex plans, read the language, not the total. A downshift, a "disciplined pace," a pivot toward lighter-asset or partner-financed models — that is Reading One being confirmed by the people with the best demand data on earth. If guidance instead holds and accelerates, the ROI warning is premature and the capex cycle still has legs. This is the single most important variable, and it resets every quarter.
The second is the open-versus-closed benchmark gap. Track the public leaderboards — coding, reasoning, long-context. Every time an open-weight model closes the gap to the frontier, the pricing power of the closed stack compresses and the ROI math for the infrastructure beneath it worsens. When open-weight models enter the top tier outright, the inference price war becomes structural rather than cyclical, and no amount of volume rescues the incumbents.
The falsification condition for the crypto side is more specific and more damning. Strip the token emissions out of any decentralized compute network and measure what is left: paid revenue, active utilization, and provider retention. If a network shows real paying utilization that survives the removal of the subsidy, it is a business. If it does not, it was a subsidy wearing a business's clothes, and its token price was never a price — it was a countdown.
Takeaway: Watch the Two Numbers, Not the Headline
The ambiguity in "token prices" will not be resolved by argument. It will be resolved by two public numbers, and then the headline will look like exactly what it was — a Rorschach blot that let every reader see their own position.
What I want you to carry forward is narrower and colder than either reading. The demand for AI is not in question. The return on AI capital is. And in 2026, in a market that has already repriced leverage, speculation, and narrative premiums, the only positions that hold are those built on the scarcest inputs and the surviving balance sheets. Two ledgers were hidden inside one word. Only one of them will still be solvent when this cycle finishes. The bank did not tell you which. The crypto outlet did not either. That silence is not an oversight — it is the trade, and it is still open. Ledger update: follow the money to the ledger that has to survive, not the one that got the headline.