- AI cost discussions are shifting from token prices to measuring the cost per completed unit of work, with attribution to business units becoming the new standard.
- Enterprises are adopting tokenomics frameworks to manage rising inference costs, as agentic AI and production workloads drive consumption faster than falling unit prices.
- Companies like Liqid are restructuring leadership to capitalize on memory and GPU pooling, aiming to deliver superior tokens per dollar, watt, and second.
The conversation around artificial intelligence costs has taken a decisive turn. For years, the debate centered on what a token costs and whether prices were falling. That argument has largely settled, and the industry has landed on a more meaningful metric: the cost of getting a unit of work done. A hefty token bill is acceptable if the overall workflow is cheaper than before, while a small bill is still wasteful if nothing useful comes out the other end. This shift marks the beginning of a harder problem—figuring out whose cost it actually is.
As enterprises move from experimentation to large-scale deployment, the spotlight has turned to tokenomics, the economic framework for measuring and managing AI consumption. This new calculus is not just about tracking expenses; it’s about attributing costs to specific business units, understanding the trade-offs between visible compute bills and invisible labor costs, and making architecture decisions that set the long-term cost base. The companies getting this right are those that treat tokenomics as a strategic discipline, not just a budgeting exercise.
The Attribution Imperative: Moving Beyond Spending Caps
When AI bills grow, the instinct is often to cap spending. Yet, most enterprises that try this approach end up with a queue that becomes a workaround, and the spend reappears where nobody is looking. The organizations that are getting it right are doing something less intuitive: they tag every token at the point of issuance, tie it to the person or system consuming it, and charge it to the business unit that owns them. If a business wants to spend more, it finds the offset inside its own cost base. Daily limits still exist, but as soft alerts rather than hard stops, because the purpose is to start a conversation rather than interrupt someone’s work.
This unglamorous plumbing is the precondition for everything else. Without attribution, cost control is theater, because no one can be held to a number they cannot see. With it, the conversation stops being about whether AI is expensive and starts being about which parts of the business are getting value from it. The spending cap, by contrast, is a blunt instrument that often fails to address the root cause of runaway costs.
The Production Handover: When Cost Responsibility Shifts
The second discipline is knowing when the bill changes hands. Central technology funds the experiments, the platform, and the build, but the business assumes the run cost the day the capability goes live. That transfer changes behavior immediately, because someone now inherits a bill they did not previously carry. It makes the promotion decision real and does more for spending discipline than any policy written on paper.
It also answers a question engineers keep asking: when should they start caring about cost? The answer is not during experimentation. During that phase, speed beats efficiency, and optimizing early kills good ideas before anyone learns whether they work. Once something is in production, the priority inverts, and the effort goes into bringing cost down while holding quality. The gains at that stage come from routing, prompt design, and workflow structure rather than from renegotiating price, and they are large enough to be treated as a standing engineering commitment rather than a cleanup exercise.
The Hidden Costs: Wrappers, Code, and Unmodeled Curves
The surprises in AI cost are never the obvious ones. The first is the wrapper. A security and control overlay can double the token cost of the workflow it was protecting. That is not an argument against the control; it is an argument for knowing what share of AI run cost sits in security before a board asks. No benchmark for that number exists today.
The second is your own code. Messy systems raise token burn because the model spends its effort working out what your systems do before it can do anything useful. Clean architecture has spent thirty years as an engineering argument that was difficult to fund. It is now a cost argument, and that one wins. The third has not fully arrived. As agents call other agents, and as regulated decisions require a trail somebody can reconstruct, volume grows on a curve nobody has modeled. It should be in the forecast as a named unknown rather than absent from it.
Falling Unit Prices vs. Rising Bills: A Forecasting Challenge
Unit prices are falling, sometimes sharply. That is true and it is the fact most often cited to close the conversation. It does not address the first problem, which is that consumption is growing faster than price is dropping, so the bill rises in a falling market. That is a forecasting problem, and attribution is most of the answer, because nothing can be forecast that cannot be seen by business unit.
The second problem is the unit price itself, and this one is a genuine exposure. The major providers are pricing into a land grab, ahead of profitability, against very large capital commitments. Whether that is deliberate subsidy or real scale economics is not knowable from outside. What follows either way is that today’s price is a poor planning assumption. Run the business case at a multiple of it and treat anything that fails as having a shelf life. The multiple matters less than the discipline of having one.
Labor as a Token Cost: The Invisible Exchange
Here is the reframe that changes the way you might approach this. Think about what it actually costs to produce enterprise software. Requirements argued in meetings, specifications written in documents, review comments in email, design debated for an afternoon. That is people generating language, it has always been the largest part of the cost, and no company has ever measured it, because it sits inside salary.
Compute tokens to do the same work are metered, itemized, and delivered on an invoice. So one of these two costs is fully visible and the other never has been, which is the entire reason a compute bill that is small next to engineering payroll still feels alarming. Do the comparison properly and in most cases the compute cost of producing an output sits well below the labor cost it displaces. What matters then is the rate at which one is exchanged for the other, and that rate is set by architecture. Routing, orchestration, model selection, and code quality all determine how much compute is burned to produce the same result. Get them right and the trade is a bargain. Get them wrong and it is a poor trade at any token price.
Industry Moves: Liqid’s Restructuring and the Memory Pooling Opportunity
Composable systems software supplier Liqid has refreshed and strengthened its senior executive team as AI tokenomics brings memory pooling front and centre to accelerate GPUs. Founded in 2015, Liqid has raised a total of $150 million, with the last round being a $100 million C-round in December 2021. It has been selling and developing its Matrix composable systems software for twenty years. The software composes virtual servers, using pools of separate CPU+DRAM, GPU, Optane SSDs, FPGA, networking, and storage resources accessed over a PCI fabric.
Since then, Optane has gone away, CXL has become a stronger offering, and as memory pools exist on CXL fabrics, composability software will be needed to dynamically allocate it to servers. Liqid launched a reference architecture at SC23 to build a composable system with a Dell PowerEdge server fitted with up to 16 Nvidia GPUs. This focus on memory pooling and GPUs was farsighted because ChatGPT started the AI large language model boom at the end of 2022, and Liqid found itself ideally positioned to help memory-constrained GPU server systems.
In July 2025, it unveiled composable GPU servers with CXL 2.0 memory pooling, announcing products enabling host server apps to access dynamically orchestrated GPU server systems built from pools of GPU, memory, and storage, focused on AI inferencing and agents. This year we have seen the rise in Nvidia-driven KV Caching schemes and a focus on tokens. The faster GPUs can process them the better, and the key is memory. The more memory a GPU server can access the better it can do its work, but HBM has limited capacity. KV caching schemes are a fix, but every off-HBM data access costs latency wasted time. Better, Liqid says, to have dynamically constructed pools of memory shared by GPU servers.
Liqid realized that the market opportunity was broader and deeper, and expanding faster than it could deal with, as Neoclouds joined enterprises and OEMs as sales targets. The board, including CEO Edgar Masri, set about transforming the company to be more aggressive in chasing this AI-boosted, memory-focused, composability market. Liqid has, in nine months, replaced its CEO, CFO, CRO, Chief Product Officer, hired a Chief Commercial Officer, and appointed a board-level, growth-focused, heavy hitter. The timeline includes Dave Larson as interim CEO, AJ Dye as CFO, Rick Hegberg as CEO, co-founder J Scott Cannata returning as Senior Fellow, Tim Pitcher as CRO, John Byrne joining the board, Dan Berg as Chief Product Officer, and Eric Shiroke as Chief Commercial Officer.
When Liqid appointed John Byrne to its board in May this year, it said a “global sales, channel, and OEM go-to-market leader joins Liqid’s Board as the company scales to deliver superior tokenomics across enterprise organizations, Neoclouds, and cloud service providers.” Byrne has more than three decades of experience at Dell Technologies, Dell EMC, and AMD. Liqid said: “AI is at an inflection point. Enterprises racing to deploy AI inferencing are confronting a memory wall and stranded GPU capacity that legacy server architectures cannot solve. Liqid’s memory and GPU pooling, powered by Liqid Matrix software, is being adopted across major OEM, enterprise, and service provider customers to increase performance and optimize utilization to deliver superior tokens per dollar, tokens per watt, and tokens per second.”
CEO Rick Hegberg said: “Liqid is growing because our memory and GPU pooling is becoming foundational to delivering superior tokenomics. John’s leadership and counsel will help us achieve broader OEM reach, deeper customer impact, and the kind of operating discipline that will help Liqid reach a new plateau in this evolving market.” Byrne said: “Memory and GPU pooling infrastructure is no longer a future-state idea; it is proving to be the way enterprises will deploy AI at scale. Liqid has all the components for accelerated growth: an exceptional team, the deepest patent portfolio in the space, and a software-defined approach that fits how the world’s leading OEMs and customers actually buy.”
Market Dynamics: Uniswap’s Tokenomics Under Scrutiny
In the crypto world, Uniswap (UNI) experienced a 5.04% decline over the last 24 hours, a move that appears driven by a combination of market sentiment, positioning, and growing skepticism about UNI’s tokenomics rather than a single discrete event. UNI’s drop significantly outpaced the overall crypto market, which was only marginally down by around 0.3% in the same period. This suggests that the decline is more UNI-specific rather than a broad market shock.
Recent discussions on X have highlighted concerns about UNI’s tokenomics and the impact of the recent fee switch on Uniswap’s pools. One widely shared post argues that the fee switch is unsustainable for Uniswap’s competitiveness, claiming that volume shares for affected pools have significantly declined. Another comparison thread contrasts UNI with Hyperliquid’s HYPE token, noting that while Uniswap generated about $99.06 million in fees over the last 30 days, only around 4.6% of that reached UNI holders, compared to about 70% for HYPE. These discussions contribute to a narrative that UNI’s tokenomics may be weaker than those of newer DeFi tokens, which could pressure the token’s price in a cautious market.
UNI’s recent price history also plays a role in its current decline. There is evidence that UNI had been repriced higher based on optimistic long-term targets, such as those from Standard Chartered. As market sentiment cools and the focus shifts back to UNI’s token-level economics, some traders may be unwinding or shorting their positions. Social media posts reference earlier bullish calls somewhat tongue-in-cheek, suggesting that the market is reassessing the realism of those projections.
Board Questions and the Path Forward
So what should a board actually ask? Not how much the enterprise is spending on AI, which is the question most are asking and the least informative one available. The better ones are: can we state the cost of our largest AI workflows per completed transaction, and if not, what is stopping us? Is every token attributable to a business unit that owns the outcome? Which of our deployments survive at a materially higher unit price? What share of our run cost is the control overlay, and is that number defensible? And when the architecture decisions that set our cost base get made, is anyone from finance in the room?
That last one is the one that matters most, because the trade at the center of this can only be booked by finance. Accepting a visible metered cost in exchange for reducing an invisible one is not a decision a technology leader can make alone. The saving lands in headcount and the cost lands on an invoice, and only a CFO can treat those as the same transaction. The enterprises that lead in this might spend more on tokens than their peers and less on getting the work done. The rest will keep managing the only number they can see.
As the industry moves forward, the pattern is clear: tokenomics is not just a buzzword but a fundamental shift in how AI costs are understood and managed. From Liqid’s restructuring to Uniswap’s market challenges, the common thread is the need for transparent, attributable, and strategically managed AI spending. The companies that embrace this discipline will be better positioned to harness AI’s potential without being blindsided by its costs.