The Tokenomics Foundation announced itself this week. The announcement was a press release. The press release ran on Crypto Briefing, not on a foundation website, because there is no foundation website. There is no list of founding members. No charter. No draft specification. No reference implementation. No open-source repository. No governance structure. No budget disclosure. No venue for public technical comment. The foundation says its mission is to standardize AI token measurement. It says the effort will shape enterprise cost management and AI investment strategies. And it says, more than once, that it has nothing to do with crypto.
The word 'tokenomics' was born inside crypto. It is a portmanteau of token and economics, forged in the 2017 ICO fire and refined by the 2021 DeFi mania. It describes supply schedules, emission curves, staking yields, and the general art of making a digital asset look scarce on paper. The Tokenomics Foundation is aware of the word's lineage. The disclaimers exist precisely because the name drags that baggage. But after three market cycles and one partial exchange freeze, I have learned that denials are never explanations. They are tells.
The code doesn't lie. People do. And in this case, there isn't even code. In 2017, I spent six weeks reverse-engineering the bonding curve logic of an AMM prototype in Chengdu, auditing the smart contract before its token launch. I found three critical integer overflow vulnerabilities that could have drained the pool. The whitepaper never mentioned them. That experience wired me permanently: verify through mechanics, not through narrative. The Tokenomics Foundation has no mechanics to verify. It has a name, a press release, and a disclaimer.
None of this means the problem the foundation names is imaginary. AI token measurement is genuinely broken, and enterprises are bleeding money because of it. The story here is not that a standard is coming. The story is that the gap is wide open, and this foundation is a signal — possibly a genuine attempt, possibly a PR game — that somebody noticed.
Context: The Enterprise AI Cost Problem Is Measured in Counterfeit Units
Let's start with the pain.
Every company that has put LLM APIs into production has hit the same wall: the bill is denominated in tokens, and tokens are not a uniform unit of measurement. A token is whatever the provider's tokenizer says it is. OpenAI's GPT-4o family uses a Byte Pair Encoding variant implemented in tiktoken, with roughly 100,256 vocabulary entries. Anthropic's Claude models ship a SentencePiece-style tokenizer with a vocabulary in a similar range but different merge rules. Meta's Llama family uses byte-level BPE tokenizers with vocabulary sizes up to 128,000. Google's Gemini has yet another vocabulary, optimized for different language coverage.
Each of these tokenizers converts raw text into subword units. They are all called tokens. They are not comparable.
Now scale that to a production environment. An enterprise spends $50,000 a month on OpenAI and $30,000 a month on Anthropic, routing requests through a gateway. The finance team wants a blended cost per output unit across the AI portfolio. It cannot get one, because the denominators are not rooted in an objective reference. One vendor's token count is derived from a vocabulary tuned on its training data. The other's is tuned on different data. Both call the result a token.
The problem is amplified for non-English workloads. Take a common Chinese character. OpenAI's tokenizer might encode it as a single token; another provider's tokenizer might split it into two or three. The same prompt in Chinese can generate substantially different token bills across providers, not because of model quality but because of tokenizer geometry. There is no standard way to audit whether a token count is fair, or even roughly calibrated.
Traditional software had units with objective referents: bytes, seconds, transactions, API calls. Cloud pricing worked because metering was transparent. The AI token broke that contract. It is a unit defined by each vendor, for each model, in each API version, with no independent verification layer.
The old-school answer to this problem was the meter. Utility companies did not trust their customers to report consumption, and they did not allow each appliance maker to define the kilowatt-hour. There was a public standard, a calibrated meter, and a regulatory regime. The AI industry has none of that. The Tokenomics Foundation is proposing to be, effectively, the Bureau of Standards for AI meters. That is an audacious ambition. The release does not say how it plans to do it.
Three possible readings of the foundation's mandate exist, and they are not mutually exclusive. First, it could be trying to define a canonical unit of token measurement, a metrological definition that every vendor's tokenizer maps into. Second, it could be trying to standardize billing information, forcing vendors to disclose tokenizer behavior and metering metadata on invoices. Third, it could be trying to define a compliance and audit framework for AI cost reporting.
Each of these is a different project with different stakeholders, different difficulty levels, and different legal implications. The release does not tell us which one the foundation is pursuing. That absence of specificity is the first serious problem. Print a press release without a scope and you have not announced a standard. You have announced a committee looking for a reason to exist.
Core Part I: The Tokenizer Layer — Why a Standard Is Technically Wicked
Let me be precise about the technical core.
Byte Pair Encoding starts with a corpus and a target vocabulary size. The algorithm counts character or byte frequencies, then iteratively merges the most frequent adjacent pair into a new symbol. 't' and 'h' become 'th'. 'th' and 'e' become 'the'. It repeats until the vocabulary hits the target size. GPT-4's tokenizer runs this process on a massive multilingual corpus, yielding a vocabulary of around 100,256 tokens. Llama's tokenizer does the same on a different corpus, with a different target size and different pre-tokenization rules.
SentencePiece operates at the level of Unicode codepoints and can bypass white-space-based pre-tokenization entirely. That makes it better for languages like Japanese, Chinese, or Thai, where word boundaries are not marked by spaces. It also allows subword regularization, a training-time technique that randomizes tokenization. That means the same sentence can map to different token sequences across multiple runs of the same model family. Nobody audits for this. Nobody can.
Byte-level BPE shifts the base alphabet from characters to bytes. It is robust across all Unicode, at the cost of longer average token sequences for some scripts. Meta's Llama uses this approach. It is well-engineered. It is also different from its competitors' approaches.
Now, the consequence. Token counts for the same input vary across tokenizers in ways that are not random: they are systematic. A standard English contract clause of 200 words might tokenize to somewhere between 150 and 260 tokens, depending on the tokenizer. The spread is not small rounding noise. It is a 70% pricing divergence, baked into the unit itself.
Numbers are a nightmare. A string like '42,000.50' can be one token in one tokenizer and four in another, depending on how pre-tokenization splits punctuation and digits. Currency symbols, commas, decimal points, and leading zeros all change the split. Code is worse: 'lambda x: x + 1' tokenizes differently not only across vendors but across code languages, indentation styles, and operator spacing. The same function in Python and JavaScript can have materially different token counts under the same tokenizer, meaning the cost of serving the same logical operation is a function of syntax, not semantics.
Multilingual text makes the divergence structural. English text runs roughly four characters per token. Chinese text runs closer to one character per token, sometimes more. A tokenizer optimized for English will naturally charge more tokens for Chinese, and a tokenizer optimized for Chinese will charge more for code mixed with prose. These are not engineering defects. They are modeling trade-offs. But they become unconscionable differences in billing when APIs are priced by token.
Now add multimodal. Vision-language models convert images into tokens via patch encoders. An image might be divided into 14x14 pixel patches, each patch becoming one visual token after projection. Audio models sample and quantize waveforms into frames that are treated as tokens. Video becomes a three-dimensional multiplication of frames, patches, and time. Every vendor chooses its own patch size, its own frame rate, and its own conversion ratio between raw pixels and tokens. One provider's API may bill a standard 1024x1024 image as approximately 1,100 tokens; another's may bill it as 2,700. The image has not changed. The pixel content has not changed. The token is a pricing unit sliced to each vendor's advantage.
The deeper problem is that token count is not a proxy for compute, nor is it a proxy for value. A model with a highly optimized tokenizer can pack more semantic content into fewer tokens, producing a lower bill for the same output quality. A model with a poorly optimized tokenizer penalizes users for the model's own inefficiency. When buyers compare price-per-token across vendors, they are not comparing like with like. They are comparing two different definitions of the unit, plus two different levels of semantic density inside that unit. A measurement standard that ignores semantic density will be gamed in exactly the same way that mortgage-backed security ratings were gamed: the rating is real, the meaning is hollow.
A measurement standard for AI tokens therefore faces an impossible breadth: text tokenization, billing metering, inference throughput, multimodal conversion, cost allocation metadata, and audit conformance. Each of these is a different standards project. A foundation that treats them as one release cycle is a foundation that has not understood its own problem.
Core Part II: The Meta-Standard Problem
Here is the structural insight that matters most.
The Tokenomics Foundation is not proposing a model, a protocol, or a tool. It is proposing a meta-standard — a standard about how other standards measure and reference tokens. Meta-standards are the hardest category in standardization because they require agreement from both the producers and the consumers of the underlying measurements, and they fail when they are either too abstract to implement or too prescriptive to be adopted.
Consider the ISO definition of the second. It is a metrological standard: a precise definition of time that anyone can implement with a cesium atomic clock. Consider the meter: defined by the distance light travels in a vacuum over a fixed time interval. These work because there is an unambiguous physical referent and because a global community of metrologists maintains the definition. Tokens have no physical referent. They are artifacts of machine learning training pipelines, each tied to a particular model's vocabulary and merge history. The Tokenomics Foundation cannot define 'one token' the way the SI defines 'one second' without inventing a canonical tokenizer — a move that would degrade every model attached to a different tokenizer.
What it can do is define a canonical unit of account: a 'standard token-equivalent' derived from a reference grammar, with mappings from each vendor's tokenizer. That is a currency exchange problem, not a physics problem. It requires every vendor to publish their tokenizer's behavior, including edge cases, in a machine-readable format. No vendor has done this. Some tokenizers are open source. OpenAI's tiktoken is open. Anthropic's tokenizer is not fully published. Google's is a moving target.
A working meta-standard would also need test vectors: a corpus of inputs, in many languages and data types, with known token counts across all conformance-compliant tokenizers. That is the equivalent of cryptographic test vectors. No one has built this corpus. Building it well requires linguistic expertise across dozens of languages, not just engineering skill. Most AI teams cannot name more than three tokenizers. The Tokenomics Foundation has not named one.
Then there is the versioning problem. Tokenizers change as models are released. GPT-4o's tokenizer differs from GPT-3.5's. Llama 3's differs from Llama 2's. A standard must specify not only the unit but the versioning of tokenizer definitions, the deprecation policy, and the migration path. Without that, compliance becomes archaeology.
The release answers none of this. That is not necessarily fatal. Early standards announcements are often deliberately vague. But the combination of vagueness with total absence of artifacts is a warning pattern. In the standards world, the announcement is supposed to follow the work, not precede it. The Tokenomics Foundation has inverted that order.
Core Part III: The Landscape — A Crowded Room With No King
Let me map the existing field, because the Tokenomics Foundation is entering a landscape that is already marked by several 'soft standards' and one de facto standard.
OpenTelemetry's GenAI semantic conventions define fields for observability: prompt tokens, completion tokens, and total tokens. The conventions tell instrumentation libraries where to record token counts and how to label them. They do not define what a token is. They simply capture whatever the model provider returns. That has the effect of institutionalizing each vendor's tokenizer as canonical for its own telemetry. It makes the data uniform in shape but non-uniform in meaning. That is the opposite of a standard: it legitimizes incompatibility while making it look orderly.
FinOps Foundation has built a cost-management discipline around cloud spend. Its framework addresses compute, storage, and networking. It has started to look at AI cost but has not produced a token definition. FinOps practitioners need token-level cost allocation now, and the absence of a token standard is a pain they feel daily. They are a natural partner for any working group that actually tries to solve this. But the FinOps Foundation is an established body with real members and real processes. A new group with no members cannot easily absorb their trust.
MLCommons runs model benchmarks. It defines tokens in the context of specific benchmark workloads, tied to specific models, and does not bind API vendors. It is a floor for performance comparison, not an arbiter of billing.
Cloud providers themselves are in a peculiar position. AWS, Azure, and Google Cloud resell LLM access. They have the engineering capacity and the financial incentive to unify token definitions across their catalogs — a single standard would make their marketplaces more direct competitors to each other. Yet none has moved. The reason is that unification would expose their pricing structures to immediate comparison and compress margins.
De facto, the closest thing to a global token standard today is OpenAI's tiktoken library. Open-source projects and observability tools copy its tokenizer behavior because it is widely available, well documented, and reasonably accurate for GPT-family models. Every LLM application defaults to tiktoken when uncertain. That means the de facto standard for 'what is a token' is a corporate implementation with no formal governance, no versioning contract, and no audit body.
The Tokenomics Foundation could, theoretically, become the neutral arbiter that the ecosystem lacks. But a neutral arbiter needs neutrality, resources, and technical credibility. The foundation has not demonstrated any of the three.
Let me also mention the commercial layer. The companies that have actually built token-accounting products — observability platforms, LLM gateways, FinOps dashboards — have solved the problem in practice by building internal conversion tables, per-provider parsers, and heuristic approximations. They are the ones closest to the actual mechanics. If the foundation wants to succeed, it must bring these companies into the room. There is no evidence it has.
Core Part IV: The Incentive Trap — Why Vendors Will Smile and Do Nothing
Now the uncomfortable part.
Token standardization benefits buyers. It enables price comparison, competitive pressure, and cost allocation. It harms sellers who rely on token ambiguity as a form of pricing power. OpenAI, Anthropic, and Google each have an interest in keeping the token unit as opaque as possible. Their API products are differentiated, the metering is proprietary, and the token estimate that appears in each response is a black-box output.
Why would a vendor join a standard that makes its offerings more comparable and, therefore, more price-compressed?
The likely behavior is 'engage and neutralize.' Vendor representatives attend meetings. They praise the initiative. They explain that their tokenizer is uniquely constrained by model architecture and therefore cannot conform to a canonical unit without sacrificing quality. They offer to 'provide input.' They commit to 'monitoring' and to 'alignment.' The result is a standard with a footnote for every vendor, which then collapses into a brand exercise rather than a measured specification.
This is the classic failure mode of soft standards. A foundation with no code, no members, and no draft emerges as the coordinating body. It sells founding memberships, certificate badges, and conference sponsorship. It publishes a glossy framework document. It produces no test suite, no reference implementation, and no enforcement. That is not a standard. That is a marketing vertical.
The counterweight in traditional standard-setting was buyer power. Government procurement, enterprise procurement, and utilities forced standards into existence because they had the purchasing weight. The AI API market is still young enough that enterprise buyers are fragmented and rarely coordinate. A foundation seeking to create a standard must first create a coalition of large buyers that will require the standard in their procurement contracts. That is the only lever that moves vendors.
Has the Tokenomics Foundation signaled any buyer coalition? The release suggests an audience of enterprise buyers, CFOs, and FinOps teams. But it names none. Without named buyers, the effort is an academic exercise at best.
There is a second-order incentive problem. Even if a vendor wanted to comply, compliance carries a real engineering cost. Publishing tokenizer behavior means maintaining documentation, supporting conformance tooling, and accepting third-party audits. That is a long-term commitment. A standards organization built on a press release cannot demand that commitment from anyone.
Core Part V: Governance — What a Standard Actually Needs
Let me be concrete about what a credible standard would require.
First, a technical charter. The foundation must publish its scope: which sub-problems it will address, what a 'standard token' means metrologically, and how it will handle multimodal tokens, versioning, and edge cases. A charted scope is the first line of credibility.
Second, an open reference implementation. A token standard without code is philosophy. The foundation must ship a library that implements the canonical unit, a conformance test suite, and a corpus of test vectors covering languages, scripts, numbers, code, and multimodal inputs. This is not optional. Cryptographic standards have reference implementations. Network protocols have reference implementations. A measurement standard has to prove itself in bytes.
Third, a public governance model. Who votes? How are working groups formed? Who publishes the standard, under what license? Is it W3C-style with member organizations, Linux Foundation-style with a neutral home, or ISO-style with national bodies? Each model has trade-offs. The absence of any governance model is a disqualifying omission for a standards body.
Fourth, an audit mechanism. Standards without verification become stickers. The industry needs a way to test a vendor's claim that its token count conforms to the standard. That requires a certification process, independent auditors, and dispute resolution. In crypto, 'audit' became theater, with firms paid by the projects they assessed. The Tokenomics Foundation must avoid the same collapse into paid compliance.
Fifth, interoperability. The standard must integrate with OpenTelemetry for tracing, with cloud billing feeds, and with FinOps tools. A standard that does not plug into the existing observability stack is a standard without market entry.
Look at the standards that actually survived in tech. The HTTP standard survived because it was simple, open, and implemented by everyone. ERC-20 survived in crypto because it was minimal and public. GDPR survived because it had regulatory teeth. Each succeeded for a different reason: simplicity, openness, or enforcement. The Tokenomics Foundation has not shown it understands which one it is pursuing.
The release mentions none of these requirements. It is possible the foundation will publish them shortly. It is equally possible that the founding team does not know what they do not know. Either way, the burden of proof is on them.
Core Part VI: The Crypto Tell
Now the title-level statement: 'not crypto.'
The term 'tokenomics' was not invented by AI engineers. It was invented by crypto economics, and it carried a specific meaning: the design of token supply, issuance, incentives, and value capture in a tokenized network. The 2017 ICO era revolved around tokenomics. The 2021 DeFi summer did too. The word is saturated with that history.
A genuinely technical AI standards organization would have chosen a different name: the Token Measurement Foundation, the AI Metering Institute, the Foundation for AI Cost Standards. It would not need to issue a disclaimer about crypto. The fact that the foundation chose 'Tokenomics' and then had to deny the crypto connection tells me the founders understand the word's crypto pedigree. It either comes from that world, or it is deliberately trading on the term's visibility while trying to disclaim its origin.
The publication venue reinforces the signal. Crypto Briefing is a crypto publication. A standards body targeting enterprise CFOs and FinOps teams does not typically launch its announcement in a crypto outlet. It releases through PR wires, tech press, or industry conferences. Launching on Crypto Briefing is a choice about audience. That audience is not enterprise procurement; it is the crypto-curious tech crowd.
Why does that matter? Because it frames the foundation's likely strategy. A crypto-native team understands tokenomics as economics of tokenized incentives. It might see a token standard for AI not as a metrological utility but as the settlement layer for a multi-agent AI economy. That could be a genuinely forward-looking vision. It could also be a way to rebadge the old token game for a new narrative.
I am not saying the foundation is a scam. I am saying that the structure of its announcement — crypto venue, crypto-coined name, crypto disclaimer — is a set of deliberate choices. Those choices give a better read on the project than the release text. The disclaimer 'we are not crypto' is the most crypto-native sentence in the entire document.
Experience: What LUNA, an NFT Floor Sweep, and an ETF Arb Taught Me About Counterparty Risk
I do not make this critique from the sidelines.
In May 2022, when TerraUSD de-pegged, I did not read the commentary. I read the mechanism. The mint-and-burn model had zero elasticity under a run. Within hours I held a 10x short on LUNA futures with $30,000 in collateral. Forty-eight hours later, the position was worth $450,000. I was right about the protocol and wrong about my counterparties. A portion of the profit sat on smaller exchanges, and when those exchanges froze withdrawals, 20% of the gain evaporated.
That loss taught me a permanent rule: the counterparty is part of the thesis. A foundation announcing a standard is — in financial terms — a counterparty. If you are going to plan your enterprise AI cost strategy around a future standard, you are taking on the foundation's execution risk. You need to know who is behind it, who funds it, and who commits resources to it. None of that is disclosed.
The 2021 NFT floor sweep taught me a similar lesson at the community layer. I identified an underpriced generative art collection and swept the floor with automated bots, spending $120,000 on 150 assets. The developer abandoned the roadmap, the floor dropped 95%, and I exited at a 70% loss. The market reasoning was sound. The human psychology was not. A standard announced with no community, no members, and no code has the same shape: hype is a lever, capital is the fulcrum, and right now there is no lever.
Even my successful trades reinforce the same pattern. In 2020, I ran arbitrage between Curve and Uniswap stablecoin pools and made a 340% return in three months. That worked because the units involved — dollars and stablecoins — had stable definitions. The entire trade relied on the comparability of the two sides. In 2024, my Bitcoin ETF basis arbitrage yielded a steady 12% annualized return. That trade existed because the SEC's approval created a regulated structure with defined units and transparent pricing.
Standards create tradeable markets. Fuzzy units create rent. The Tokenomics Foundation is trying to build the first. If it succeeds, it enables a whole category of cost-optimization and arbitrage strategies that are currently impossible. If it fails, it simply adds another 'standard' to the pile that nobody reads.
Contrarian: The Standard Could Make Things Worse
Here is the counter-intuitive angle that most commentary misses.
Even if the Tokenomics Foundation executes perfectly — even if it publishes a technically sound, openly governed, multi-vendor token measurement standard — the standard could still harm its intended beneficiaries.
The reason is metric fixation. Once a canonical token becomes the procurement metric, enterprises will optimize for token minimization at the expense of outcome quality. Procurement teams will choose models based on cost per standard token and will push teams toward token-efficient models regardless of whether the resulting outputs are safer, more accurate, or faster. The same distortion happened with lines-of-code metrics in software management, with click-through rates in advertising, and with engagement time in social media. A measure that becomes a target ceases to be a good measure.
A token standard would also normalize a single axis of evaluation in an industry where quality is multidimensional. Models differ in reasoning, latency, safety, instruction following, and multilingual capability. A standard that reduces the comparison to unit cost invites the enterprise to ignore the other axes. That will produce bad procurement decisions that are fully compliant with the standard.
There is also the audit capture problem. If the standard requires vendor certification, who performs the audits? In crypto, the audit industry became a paid rubber-stamp racket. Projects hired known firms, got 'passes' that meant little, and displayed the badges with pride. This is not a distant risk; it is the statistically dominant outcome of any voluntary certification ecosystem without real enforcement.
Finally, standardization can freeze the measurement layer while the underlying technology evolves. Tokenizers change with every new model generation. Context windows expand. Multimodal encoding changes. A standard that is too rigid becomes a drag on innovation. A standard that is too flexible becomes meaningless. The correct design is versioned, modular, and explicitly iterative. That is a lot harder than producing a framework document.
There is another alternative nobody mentions: private measurement. A sophisticated enterprise can build its own internal conversion layer, standardizing tokens for its own procurement decisions using open-source tokenizers and independent benchmarks. That approach does not require a foundation, a governance model, or a certification regime. It requires engineering discipline and a transparent internal ledger. For many companies, that is the rational path, not waiting for a public standard to materialize. The public standard, if it ever arrives, will be useful primarily as a shared reference. But the enterprise that builds its own metrology now will have a procurement edge that no committee can create.
So even a successful Tokenomics Foundation presents a double-edged outcome. The problem is real. The solution, if executed poorly, becomes a new form of opacity, dressed as clarity.
Takeaway: The Checklist, the Signal, and the Standard That Does Not Exist Yet
If the Tokenomics Foundation wants to be taken seriously, it can answer six questions.
One. Who are the founding members, and what is their track record in actual standards work? Names, not logos.
Two. What is the technical charter? Which sub-problem comes first: tokenizer definitions, billing metering, observability, or multimodal conversion?
Three. Where is the reference implementation and the test-vector corpus? Standards live in code, not in press releases.
Four. What is the governance model? Who votes, who publishes, who arbitrates, under what license?
Five. Who funds the work? If the backers are crypto-related, disclose it. The denial is doing more damage than the disclosure would.
Six. Which named enterprises are committed to requiring this standard in procurement? That is the only force that will make vendors comply.
No standard has ever been created by one announcement. Standards are built in working groups, test suites, and multi-year implementation cycles. The Tokenomics Foundation is at the very beginning of that path, and it has not yet shown it understands the path exists.
The market signal is still worth watching. If cloud vendors and model providers assign serious engineering time to token standardization, it moves. If the foundation acquires real members and produces a test suite within the next year, it becomes a live project. If, on the other hand, this remains a press release and a domain name, it was a ghost at the moment of its own birth — and there is no liquidity in a ghost.
I count tokens myself. I run my own benchmarks. I build internal conversion layers before I trust any vendor's meter. You should do the same.
Volatility is just interest for the impatient. Standardization is the opposite: it is the patient extraction of value from chaos. The Tokenomics Foundation is a rumor of patience. It has not yet demonstrated any.