Ask three AI vendors what a token costs and you will get three answers. Ask them what a token is, and you will get three more. For the past eighteen months, this has been the dirty secret hiding behind every glossy "cost per million tokens" price page, every enterprise AI procurement deck, and every CFO who nodded confidently at a number they could not independently verify. The unit that powers the entire generative AI economy is, in a word, unaccountable.
Here is a concrete version of the absurdity. Suppose a company feeds the same ten thousand support tickets into three different LLM APIs, asking each to generate summaries. The first vendor reports 12.4 million tokens consumed. The second reports 9.8 million. The third reports 17.2 million. The invoices are based on three different counts of the exact same input. The CFO asks which vendor is cheapest per token. The honest answer is that the question is unanswerable until someone defines what a token actually is.
Last week, an organization calling itself the Tokenomics Foundation announced its existence. Its stated mission: standardize how AI tokens are measured so that enterprises can finally compare costs across model providers, cloud platforms, and API vendors. Its secondary mission, repeated with noticeable urgency: this has nothing to do with cryptocurrency. The name, the foundation insists, is simply what happens when the word "tokenomics" migrates from crypto's whitepaper graveyard into the AI dictionary.
I have spent enough years in digital assets to know that nothing with the word "tokenomics" in it is accidental. The term was born in the crypto ecosystem, popularized by a thousand ICO whitepapers, and weaponized by communities that measured value in decimals and faith. A foundation choosing that name while loudly denying any crypto connection is either navigating a compliance minefield or courting the very attention it claims to reject. Both explanations are worth watching. Neither tells us whether the underlying project is sound.
Here is the part that interests me as an economist rather than a skeptic: the Tokenomics Foundation, as far as anyone can tell, has no website, no published membership list, no founding partners, no budget disclosure, and no draft standard. By every observable measure, it is a press release with an ambition attached. That should make it easy to dismiss. It should not. Because the problem this foundation claims to solve is real, it is urgent, and it is growing more expensive by the quarter.
Let me set the macro scene. When I think about the liquidity map of this moment, I see three pools. There is the institutional pool, still rotating cautiously into AI infrastructure through public equities and private credit. There is the venture pool, pouring money into model startups with stunning speed and equally stunning burn rates. And there is the enterprise operating pool — the monthly invoices that flow from businesses to API providers, already enormous and growing at a pace that taxonomies cannot keep up with. Every economic cycle has one unit that becomes the canonical measure of value. In the 2000s it was the server rack. In the 2010s it was the cloud instance-hour. In this cycle, it is the AI token. And every canonical unit, once it becomes canonical, gets standardized — usually in fits and starts, with pressure from the people who pay for it.
The average mid-size company running LLM-based workflows burns through millions of tokens per day — tokens for chat completions, for embeddings, for fine-tuning, for image generation, for audio transcription. Yet the unit itself is undefined, unregulated, and effectively unverifiable by the end customer. You can no more audit your AI invoice than you can audit the number of bytes in a compressed file by looking at its price tag. This is the vacuum the Tokenomics Foundation hopes to occupy. Vacuums attract standards bodies, metrics vendors, consultants, and grifters. The art of reading an announcement like this is deciding which category the founders belong to before their first product ever ships.
Let me pull from my own history here, because it shapes how I read every standards story. In late 2017, I was auditing early utility tokens, and the highest-signal work involved sitting in Telegram groups watching sentiment move in real time. When the Status Network ICO captured the community's imagination, token vesting schedules and liquidity anxiety were the two topics that dominated every channel. I organized town halls for more than 500 retail investors to walk through the economic model in plain language. The lesson I carried into the next cycle: a network is only as stable as its shared understanding of what it is measuring. When people cannot agree on what a unit is, they default to speculation. AI token measurement has the same structural weakness today. There is no shared understanding of the base unit, so the market defaults to whatever the vendor prints on the invoice.
Why a Token Is So Hard to Count
Let me break down why "token" is such a slippery unit, because the technical reality matters more than any press release.
Start with tokenizer disagreement. The token count for any given text depends entirely on which tokenizer splits it. OpenAI uses its own BPE-derived tokenizer. Anthropic ships a different variant. Google's Gemini family uses another. The open-source ecosystem leans on SentencePiece and byte-level encodings. Feed the sentence "I love New York City" into three different tokenizers and you will get three different token counts, three different word-piece alignments, and three different cost implications. A million tokens on one platform is simply not a million tokens on another. The input text is identical. The measured unit is not.
Let me make this concrete with an example I have actually seen in contracts. The word "tokenomics" — fittingly — is rendered as a single token by some tokenizers and as two or even three tokens by others, depending on whether the system has seen the word often enough to allocate it a dedicated token ID. Chinese and Japanese text compounds the problem: the same sentence can produce wildly different counts, often differing by thirty percent or more across providers. This is not a rounding error. A thirty percent measurement gap on a six-figure monthly AI bill is a hundred thousand dollars a quarter of phantom variance.
Second, the multimodal era has made the fragmentation systematic. A token is no longer a word fragment. Images are converted into patches. Audio becomes spectrogram frames. Video becomes spatio-temporal slices. Every vendor converts these inputs into "tokens" using proprietary formulas that serve their billing logic more than your accounting logic. GPT-4o's vision token is not Claude's vision token, which is not Gemini's. When a procurement officer asks how many tokens it costs to process a page of invoices, the honest answer is: it depends on which miracle of compression the vendor decides to apply that quarter.
Third, there are at least three different kinds of token in play at the same time. There is the inference token — the unit a model actually processes during generation. There is the billing token — the unit an API provider charges for, which may or may not include cached inputs, system prompts, tool outputs, and serialization overhead. And there is the throughput token — the tokens-per-second figure that appears in every hardware benchmark, pitting one GPU vendor against another. A standard that conflates these three is worse than no standard at all, because it gives everyone false confidence.
Then there is the pricing layer on top of the counting layer. Input tokens are cheaper than output tokens. Cached tokens are cheaper than uncached. Batch APIs are discounted further. Context-window pricing shifts as your conversation grows. Some providers change prices quarterly, some give enterprise customers custom rates that are confidential. If the Tokenomics Foundation wants a standard that makes comparison shopping possible, it is not designing a counting algorithm. It is designing a billing disclosure format that vendors will have to adopt — and vendors have precisely zero incentive to adopt it voluntarily.
I have watched this dynamic play out in another market with another kind of token. In the 2020 DeFi summer, I directed a fund allocating $2 million into Aave and Compound liquidity pools. What struck me, beyond the yield curves, was how quickly capital flowed toward protocols with legible interfaces and away from those with hidden costs. My team spent disproportionate time analyzing user experience friction points flagged in community forums — an ambiguous liquidation threshold, a confusing rate display, a poorly labeled governance vote. We coordinated with product teams to smooth those interfaces for non-technical users. The result was capital retention that outperformed nearly every pure yield strategy we tracked. The lesson has not aged: in a market where measurement is unclear, users read clarity as a proxy for trust. The same logic now applies to enterprise AI buyers. They will read token-billing clarity as a proxy for vendor integrity.
The Room Is Already Full
The existing competitive landscape makes the Tokenomics Foundation's job harder, not easier. OpenTelemetry's GenAI semantic conventions already define observability fields for token usage, latency, and cost metadata. Those conventions are not a pricing standard, but they create the technological substrate that any future pricing standard will need to align with. The FinOps Foundation has spent years building a vocabulary for cloud cost management and is actively folding AI spend into its frameworks. MLCommons already benchmarks model quality through MLPerf and related harnesses; adding a "cost per million output tokens" benchmark would be trivial.
And then there are the commercial tooling players — Datadog, LangSmith, Helicone, and a dozen observability startups — who already sit on token telemetry. They have the logs, the dashboards, and the customer relationships. If any of them ships a standardized token accounting layer that wraps every major API, they do not need a foundation's permission to define a de facto standard. They need a press release about ten percent as ambitious as this one — and the network effects will do the rest. This is the uncomfortable truth for any standards foundation: in the age of software, the team that owns the instrumentation owns the standard. Formal bodies and working groups are increasingly appendices to the real standard-setters, which are the tools people run in production.
A standards organization entering this field without members, funding, or a reference implementation is not entering a vacuum. It is entering a crowded room and hoping the existing occupants will move aside. History suggests they will not — unless the new entrant brings a constituency the incumbents need.
Why Sellers Hate Clear Meters
This is where the economics gets uncomfortable, and it is the part most coverage of this announcement will avoid. Standards are not adopted because they are better. Standards are adopted because someone with buying power forces the issue, or because a monopolist decides to gift a de facto standard to the ecosystem for strategic reasons. USB came from a consortium of buyers — Intel, Microsoft, Compaq — that had enough market power to coerce the peripherals industry. HTTP came from the web's distributed need for a common language. Shipping containers came from trucking and rail companies that could no longer tolerate the handling costs of break-bulk cargo. In every case, the standard was an act of convenience for the parties that held the real leverage.
For AI tokens, the parties holding real leverage are the model providers: OpenAI, Anthropic, Google, and the hyperscalers that distribute their models. They benefit from pricing opacity because opacity raises switching costs and allows subtle bundling. A buyer-friendly token standard would force them to expose their accounting — roughly as attractive to them as a calorie-counting law is to a fast-food chain famous for secret sauces. This is the core paradox of the Tokenomics Foundation's mission. To succeed, it needs the cooperation of the very actors whose strategic advantage it threatens. The only way that happens is if buyer pressure becomes so intense that vendors conclude a transparent standard is the least-bad outcome.
We are approaching that threshold. Enterprise FinOps teams are starting to demand breakdowns, and the demand is loud enough that vendors are beginning to accede, one spreadsheet at a time. This is the window the foundation is trying to hit: a moment when purchasers are angry enough to want answers but not yet organized enough to demand them. History repeats, but liquidity decides the tempo. Right now the liquidity is in AI — hundreds of billions of dollars of enterprise spend flowing through APIs every quarter — and where liquidity concentrates, standards eventually follow.
The measurement will also reshape cost structures in ways nobody is forecasting. Look at what happened to Ethereum's Layer2 ecosystem after the Dencun upgrade made blob data cheap overnight. Every rollup budgeted for that era of low fees. But blob space is finite, and within two years saturation will push rollup gas fees back up — and every team that built a cost model on the cheap era will feel the pain. Token measurement standards operate the same way. The cost structure you measure today is not the cost structure you will pay tomorrow. A standard that locks in today's token definitions will create winners and losers when pricing models shift, and if the standard is captured by vendors, it will shift in their favor.
My 2024 experience advising institutional clients on Bitcoin ETF structures reinforced this. The hard work was not the financial engineering. It was translating complex regulatory frameworks into accessible user-benefit narratives for traditional finance executives who had never touched a private key. We collaborated with legal teams and community advocates to ensure the ETF structure addressed retail accessibility, not just institutional compliance. That dual-focused bridge — technical detail translated into user adoption — is exactly what a token measurement standard requires. Commodity markets need a neutral meter. Without one, every buyer is flying blind, and every seller is selling weather forecasts.
There is also a valuation angle worth naming, because I manage capital and I think in these terms. If the Tokenomics Foundation were a company, it would be a pre-product infrastructure startup with no revenue, no team disclosure, and no clear path to monetization. Standard-setting organizations typically monetize indirectly — through certification fees, through sponsored research, through the prestige that converts into advisory mandates. But the strategic value is real. A standard that gets cited in enterprise procurement contracts becomes a toll road for every transaction in the ecosystem. That is why large cloud providers and observability companies sometimes acquire standards bodies or their key authors. The asset is not the standard itself; it is the gravitational pull it exerts on purchasing decisions.
What a Real Standard Would Look Like
If I were running a diligence process on the Tokenomics Foundation — the same process I ran on utility tokens in 2017 and ETF sponsors in 2024 — I would look for four artifacts. First, a reference implementation: open-source code that counts tokens according to a published algorithm, tested across multiple models and adversarial inputs. Second, a test corpus: a public set of texts spanning languages, code, images, audio, and video, with published expected token counts from every major model. Third, a governance model: independent board, balanced stakeholder representation, a clear neutral-party arbitration mechanism for disputes. Fourth, a certification process: a way for API providers to prove their invoices match the standard's counting rules, with real consequences for non-compliance.
Think about how the electrical meter became trustworthy. It is not enough for a utility to promise accuracy; the meter must be sealed, periodically tested, and subject to public oversight, with independent audits and penalties for tampering. AI token accounting needs the same institutional machinery. A certification mark on an API provider's invoice — backed by verifiable code, spot audits, and escalating consequences for mismatch — is worth more than a thousand white papers. Without enforcement, a standard is a suggestion. With enforcement, it is a market boundary. Every serious standards story in history is ultimately an enforcement story.
The announcement offers none of these artifacts. It offers the word "standardize" and a name. That is not enough to build trust, and trust is the only asset a standards body actually has. I learned this the hard way in 2022, when Terra and Luna collapsed and I chose not to liquidate my book and go quiet. Instead I launched a "Transparent Risk" newsletter series for my 10,000 subscribers, detailing our fund's exact exposure, our hedging logic, and the positions we were unwinding. It was counterintuitive — in a crisis, most managers shut down — but it retained 85% of our capital through the worst downturn. The community treated transparency as an anchor, not a weakness. The lesson: trust is built by sharing the machinery of measurement, not just the results.
The Danger of a Pseudo-Standard
Here is the contrarian angle that most coverage of this story will miss. In the short term, the Tokenomics Foundation does not need to succeed for its announcement to have an effect. The act of declaring a standard — even an empty one — normalizes the idea that token pricing should be standardized. It opens the door for more credible actors to step through. The earliest ICO whitepapers, however flawed, educated a generation about utility tokens. This press release might be doing the intellectual dirty work that a more serious standards body will eventually exploit.
But there is a darker reading, and it is the one I cannot shake. A pseudo-standard is more dangerous than no standard. If any organization produces a measurement framework that gets adopted without genuine technical verification, it creates a false sense of comparability. Procurement teams anchor decisions to a number that looks authoritative but is not. Regulators cite it in guidance. Budgets allocate on its basis. And when the mismatch between the standard and reality surfaces — it always does — the market pays a transparency tax that dwarfs whatever billing opacity originally cost.
The complexity trap is just as real. When Uniswap V4 shipped its hooks architecture, it turned the DEX into programmable Lego; the power was undeniable, but the complexity spike scared off the vast majority of developers who tried to build with it. A token standard that tries to solve every problem at once — text, multimodal, billing, throughput, observability — will face the same fate. The widest standard is the shallowest. The standard that wins will be narrow, brutal, and honest about its limits.
And here is the blind spot I hope the foundation's founders are staring at right now: token price is not model quality. A standard that makes unit costs beautifully transparent will tempt procurement teams to optimize for the cheapest token, ignoring latency, safety, context quality, and the subtle ways models fail. Measurement is never neutral. The choice of what to count shapes what gets optimized, and what gets optimized shapes what gets valued. If token price becomes the universal procurement metric, we will collectively discover, too late, that the cheapest token sometimes comes with the most expensive consequences. A standard that ignores those tradeoffs is not measurement. It is misdirection — and so far, the foundation has said nothing about how it would prevent its own metric from becoming a weapon of cost-cutting myopia.
I have also seen how cultural narrative drives value in ways no purely financial model can capture. In 2021, my $500,000 Art Blocks portfolio outperformed because we curated around community ownership and social bonds, not speculation — the strategy returned three times over while the hype cycle collapsed around it. That is the kind of value that persists. I watched Bitcoin — the original peer-to-peer cash of Satoshi's vision — get wrapped in an ETF and reframed as a Wall Street toy. The unit survived. The vision dissolved. The same thing will happen to AI tokens if measurement becomes a marketing feature rather than a public utility. Culture is the code that compels human adoption, and the culture around AI procurement is currently one of opacity, hero worship, and spreadsheet FOMO. A standard will not fix that culture on its own. It will only be as honest as the community that enforces it.
The Ninety-Day Test
So where does this leave us? The Tokenomics Foundation is, at this moment, a name attached to a real problem. The problem is the unmeasured measurement of AI's most important unit. The name may or may not become attached to the solution. But I can tell you with confidence that AI token standardization is coming. Capital demands comparability, and history repeats, but liquidity decides the tempo.
The near-term signal to watch is not the foundation's website. It is the list of who signs on. If a single model provider, hyperscaler, or Fortune 500 FinOps team joins within the next ninety days, this becomes real. If the foundation remains alone in the press-release ether, it will be a footnote at a conference no one remembers.
Until then, the healthy response from the enterprise community is neither skepticism nor enthusiasm. It is diligent skepticism. Demand the code. Demand the test corpus. Demand the governance. If the Tokenomics Foundation cannot deliver, the door stays open for someone else — a tooling vendor, a benchmark consortium, or a coalition of angry CFOs — to build the meter that AI's economy so desperately needs. A meter will be built. The only open question is who reads it, and whether the reading serves buyers or sellers. That is not a technical question. It is a community question. And communities, as always, get the standards they demand.