How to build an AI-ready sovereign enterprise
As AI spreads through the enterprise, companies are losing something that few have thought to protect: the capacity to judge what their systems are doing and to change course when they need...
Artificial Intelligence
AI keeps getting cheaper. Your bill keeps going up
AI is developing âmodel-yearâ pricing: old capabilities get cheaper while frontier models stay expensive. Michael Watkins explains what this means for competitive advantage.
Rapid read:
- Older AI capabilities are rapidly becoming cheaper, while the latest frontier capabilities remain expensive because they require more compute and offer the newest forms of performance.Â
- Commoditized AI cannot create durable competitive advantage, so executives must distinguish between capabilities that will soon be available to everyone and those that can still differentiate the business.Â
- Competitive advantage will come from investing selectively at the frontier, protecting complementary assets such as proprietary data, trust and distribution, and reconfiguring quickly as each model year depreciates.
Speaking at BlackRockâs US Infrastructure Summit in March 2026, OpenAI CEO Sam Altman described a future in which âintelligence is a utility, like electricity or water, and people buy it from us on a meter.â
With the cost for companies of scaling AI ballooning, framing future pricing as similar to utilities is comforting for business leaders. Utilities are cheap, and cheap intelligence would mean AI eventually stops being a line-item worth arguing about, in the same way that internet bandwidth and long-distance calling did: plan accordingly, wait, and the cost problem will solve itself.
Except it doesnât. And the bandwidth comparison illustrates why. The price of transmitting a megabyte of data collapsed, but corporate spending on connectivity did not. Once bandwidth got cheap, firms streamed video in addition to sending emails. The volume they consumed grew faster than the price fell. AI capacity is likely to follow a similar trajectory, but at a faster pace.
When OpenAI first sold access to GPT-3 in late 2021, processing a million tokens of text â roughly 750,000 words â cost $60. Three years later, models matching GPT-3 performance cost six cents for the same volume â a thousand-fold reduction. Yet over the same period, the bill for using the most capable current model remained high. Intelligence is not becoming uniformly cheap. It is becoming cheap with a lag. Last yearâs capability is trending toward free while this yearâs capability stays expensive. There is no clear reason that pattern wonât continue.
Understanding the two key curves
The first curve â the price of a fixed amount of AI capability â is well-documented and shows the price is falling rapidly. Epoch AI, a research group that tracks the cost and capabilities of AI, tracked six benchmarks and found that the price to reach a fixed performance level fell between nine-fold and 900-fold per year depending on the task. The price of GPT-4-level performance on PhD-level science questions fell roughly 40-fold per year. Andreessen Horowitzâs analysis puts the decline for equivalent-performance models at about 10-fold annually.
The mechanism driving the price reductions is competition, and increasingly it is Chinese competition. Open-weight models from DeepSeek, Alibaba, Zhipu, Moonshot, and Meta put a floor on pricing: once a capability can be replicated by a model, you can run on your own hardware. No vendor can charge much for it, and the floor is rising fast: by mid-2026, the best open-weight models trailed the proprietary frontier by months rather than years, the smallest gap yet measured. That floor is a strong argument for Altmanâs abundance thesis.
The second curve runs the other way. While last yearâs capability was falling toward free, the frontier price â what only the newest models can do â does not fall; it resets. OpenAIâs first frontier reasoning model launched at the same $60 per million tokens that GPT-3 had charged at its debut. The amount of computation a single task consumes is rising just as steeply. A chatbot question triggers one call to a model, while an agent doing the same nominal job plans, calls tools, checks its own work, and tries again â and on every step it re-sends everything it has read so far. Gartner puts agentic workloads at between five and 30 times the tokens of a standard chatbot exchange. A study of eight frontier models on a standard software-engineering benchmark, published in April, found agentic coding tasks running to roughly a thousand times the tokens of an equivalent chat request, with the re-sent input rather than the generated output driving most of the cost. The same task varied by up to 30-fold from one run to the next, and the models could not predict their own consumption. They systematically underestimated it.
Put the two together and the paradox dissolves. A bill is a price multiplied by a quantity. Suppose a unit of AI work cost a dollar last year and costs 10 cents today â a 90% cut, and a real one. If the job that used to take one unit now takes 50, the bill goes from one dollar to five. That is the model. Prices are falling. Volumes are rising faster.
This is why enterprise spending on models has climbed during the steepest price declines in the industryâs history. Menlo Ventures estimated that companiesâ spending on language models more than doubled in a six-month span in 2025, while research firm Gartner expects spending on AI models and platforms to reach $64.3bn in 2026. And the impact is visible in the budgets of individual companies too. Uberâs Chief Technology Officer revealed in April this year that the company had spent its full-year 2026 AI budget four months in, after an agentic coding tool spread across 5,000 engineers at $500 to $2,000 per engineer per month. Reports from companies about prices falling while costs skyrocket have become the norm.
This quarterâs expensive differentiator is next quarterâs free commodity, and something else has taken its place at the top. The line between the tiers never stops moving, and it moves in one direction.
A floor, a frontier, and a line that keeps moving
Put the curves together and you get the actual shape of the market: capability does not stay at the frontier. It descends, continuously and quite quickly, to the floor. This quarterâs expensive differentiator is next quarterâs free commodity, and something else has taken its place at the top. The line between the tiers never stops moving, and it moves in one direction.
Cars offer the nearest analogy: last yearâs model is cheap, this yearâs is not, and next year the same thing happens again. The image is useful provided you hold the calendar loosely. The measured interval is closer to a quarter than a year. An executive planning an annual refresh cycle is planning roughly three cycles too slow.
AI and digital transformation programs
Build the skills to lead in a digital world
Transform your career and business with next-generation AI and digital skills
When using AI stops being an advantage
Strategy has a standard test for whether a resource can confer competitive advantage: it must be valuable, rare, and costly to imitate. Last yearâs models pass the first test but fail the other two. Capability at the free floor is available to every competitor and every new entrant. Adoption becomes table stakes rather than differentiation. The organizations that spent 2024 and 2025 congratulating themselves on AI deployment will find the deployment itself conferred no durable advantage.
Strategists have seen this pattern before. When an innovation is valuable but easy to imitate, the profits flow not to the innovationâs users but to the owners of the complementary assets it needs to create value. For AI, those assets are proprietary data that canât be scraped, regulated licenses, distribution, physical infrastructure, institutional trust, and the capital to stay at the frontier when the frontier matters.
But the model-year dynamic adds something the classic frameworks donât account for. When the environment reprices capability every year, even a strong asset position erodes if the organization canât reconfigure around each new model year of AI capability. The durable advantage is what the literature on competitive advantage calls a dynamic capability: the routines for sensing which model year a capability belongs to, absorbing commodity capability faster than rivals, and redeploying people and capital as the floor rises. Static moats determine where you can win; reconfiguration speed determines whether you keep winning.
The model-year framing changes how executives should be making decisions in every industry.
Implications for executives
The model-year framing changes how executives should be making decisions in every industry. Healthcare is a useful example because it has all the relevant structural features: proprietary data, heavy regulation, licensed professionals, and thin margins. Consider the chief executive of a $10bn regional nonprofit health system: a dozen hospitals, a large outpatient network, and 40 AI projects competing for capital. The conventional way to sort them is by ROI. The better question to ask would be, âWhich of these will be free in 18 months?â
Last yearâs models. AI scribes that listen to patient visits and draft the clinical notes, tools that assign billing codes, triage of patient messages, automated handling of routine call-center traffic. Every industry has its version: first-pass contract review in legal, tier-one support in software, reconciliation in finance. These are valuable. In a multicenter study across six US health systems published in JAMA Network Open, from 52% to 39%. But the capability is already replicable, and the competitor will rapidly have it too. The benefits will show up in operating margins and retention, but not competitive position. So purchase the capability accordingly. Sign short-term contracts. Donât lock in todayâs prices for capability whose price is falling. Assume the floor will drop.
This yearâs models. The frontier in healthcare is applications such as appealing complex insurance-claim denials, optimizing patient flow across a hospital network, supporting diagnosis in ambiguous cases, and matching patients to clinical trials. Equivalents in other industries are multi-step deal analysis, supply-chain optimization, and novel engineering design. These require current-generation reasoning, cost real money per task, and are uneven in quality. Theyâre also where differentiated performance shows up. Fund a handful as time-boxed experiments with predefined clinical or financial endpoints, cut the ones that miss, then build the survivors into standing instrumentation rather than another round of pilots. A pilotâs answer expires when the next model year arrives.
Durable assets. In healthcare, an example is 20 years of longitudinal patient records linked to outcomes in a specific population. No lab can scrape that, and it may be difficult for competitors to replicate it. Your industryâs equivalents could be customer histories, regulated licenses, distribution relationships, or the physical network. Accountability belongs on this list too. Nobody is discharging a patient solely on a modelâs say-so, and patients canât sue a model for malpractice. The same holds for audit opinions, engineering certifications, and fiduciary legal advice. In professional services, accountability is an essential part of the product. It may prove the most durable asset of all: capability depreciates on the model-year schedule, but the willingness of regulators, courts, and patients to accept a human signature does not. These are the complementary assets in concrete form, and theyâre why an incumbent can coexist with better-capitalized rivals that have identical AI model access.
The inputs for the next several years of capability development are already committed: the chips are contracted, the power deals signed, the datacenter capital allocated.
Three scenarios
The real uncertainty is the lag: how far ahead the frontier stays, and whether the gap is compounding. The inputs for the next several years of capability development are already committed: the chips are contracted, the power deals signed, the datacenter capital allocated. Expansion of effort is not in doubt; the yield is. Three scenarios define the range of potential outcomes.
1 – Commodity convergence â everybody ends up with the same tools
Open-weight models track the frontier within 12 months, and frontier capability is only marginally better for practical work. Intelligence really does behave like a utility. Competitive advantage returns to where it always sat: operations, data, and execution.
Implication: Spend as little as you can on the frontier and deploy commodity capability as widely as you can across your company.
Signposts: Open-weight releases matching frontier benchmarks quickly; frontier vendors visibly cut prices under competitive pressure; capability gaps that look real on paper but that donât survive contact with real workflows.
2 – Compounding frontier â the best models stay meaningfully better
The gap holds at 18 months or more, and frontier models produce materially better reasoning, research, or optimization. Access to first-tier capability becomes a strategic input comparable to capital. In healthcare, large academic centers and national systems pull away from regional players who canât fund it.
Implication: Mid-size organizations should be forming purchasing consortia now, while they still have leverage.
Signposts: Capability differences showing up in studies that follow real outcomes over time, not just benchmark scores, vendors tiering pricing steeply; frontier-only capabilities that still have no open-weight equivalent after two years; the length of tasks models can complete autonomously continuing to double â a rate that METR, a nonprofit that evaluates frontier AI systems, puts at roughly seven months over 2019â2025 and closer to four months for models released since 2023.
3 – Regulated plateau â capability outruns the ability to deploy it
In healthcare, liability exposure, payer rules, FDA device pathways, and state AI statutes gate deployment regardless of what the models can do. The binding constraint is evidence and approval, not intelligence.
Implication: The scarce asset is evaluation and governance infrastructure: the ability to prove a system that is safe and effective in your population.
Signposts: How CMS and FDA treat systems that keep learning after approval; early malpractice case law on AI-assisted decisions; state-level regulatory divergence. Your industry almost certainly needs to anticipate a similar set of scenarios. Most organizations are implicitly planning for scenario one while their vendors are pricing for scenario two.
What executives should do
The following five actions are dynamic capabilities you need to build â routines organizations need to run continuously as the floor rises.
1 – Sort your project portfolio by model year, not just ROI.
Ensure tasks are routed to the right model. Commodity initiatives get bought cheaply and deployed broadly. Frontier initiatives get funded selectively, with real evaluation attached.
In practice: At the next capital review, add one column to the project list â the expected date each capability hits the commodity floor. Anything under 18 months gets bought, not built.
2 – Avoid long vendor lock-ins for capabilities that are falling in price.
If a frontier lab can only charge premium prices for a limited time, discounts in exchange for a long commitment or rights to your data are their way of buying insurance against commodification with your money. Price declines should accrue to you, not to your vendorâs margin. Be wary of labsâ attempts to shift from selling tokens to selling work. Agents, seats, and workflow integrations are little more than attempts to convert a depreciating asset â the model â into durable ones: distribution and switching costs.
In practice: Cap contracts for commodity capability at 12 months, and make exportability of prompts, workflows, and evaluation data a condition of signature.
3 – Treat proprietary data as a balance-sheet asset.
Govern it carefully and know exactly what youâre giving away and what you are getting for it. The companies that signed broad data-sharing terms for early access will find they traded their only durable moat for an 18-month head start.
In practice: Commission an inventory of the data the organization holds that no competitor can replicate, who currently has contractual access to it, and what was received in return. Few executive teams can answer the third question.
4 – Build evaluation capability in-house.
In every scenario, the ability to determine whether a model is actually better for your specific use is scarce and undersupplied. Itâs the closest thing to a no-regrets investment on the list. Its most developed form is a standing instrument rather than a series of studies: a live process wired so that agents run in shadow alongside people, both are scored against real outcomes, and each decision is promoted toward autonomy or demoted back as the evidence shifts.
In practice: Assign a small standing team to run every candidate model against a fixed set of the organizationâs real tasks â not vendor benchmarks â before any purchase or renewal, and budget for it as infrastructure, like security, not as a one-off study.
5 – Build governance, accountability, and trust capabilities.
Evaluation tells you what the models can do. Governance decides what they may do, and who answers when one is wrong. The two boundaries move on different schedules: a capability can be ready long before regulators, customers, or your own professionals will accept it. Treat that acceptance as something you build: a named owner for every autonomous decision, clear rules for pulling authority back the moment the evidence turns, and records good enough to defend a decision to a regulator or a court.
In the third scenario, this is the binding constraint. In the other two, itâs what lets you go faster than rivals without betting the franchise. The commodity floor also arrives with a jurisdiction question: the cheapest capable models are increasingly Chinese â permissively licensed and self-hostable, but with hosted services that can place your data under foreign data law, and without the compliance attestations regulated industries require. Cheap is not the same as deployable. A model cannot sign an audit, hold a license, or be sued for malpractice, so in regulated markets the labs must sell through incumbents rather than around them. That is leverage you should use.
In practice: Before adopting any low-cost model, ask where the data goes and who signs the compliance attestation. If the answers are âabroadâ and ânobody,â self-host or pay more.
The implementation process is summarized in Figure 1.
Five questions for the next leadership meeting
- Which of our AI initiatives will be free to our competitors within 18 months?
- Which contracts have us paying 2026 prices for what will be 2027 commodity capability?
- What data do we hold that no rival can replicate â and have we already signed away access to it?
- How would we know whether a new model is better for our work?
- And if one of our autonomous systems makes a costly error tomorrow, who answers for it?
A key question
Altman may be right that intelligence becomes abundant and so equivalent to a utility. The falling-cost trend is measurable. But abundance and cheapness are not the same as accessible-when-it-matters, and the utility analogy therefore has limited analytical power. Utilities are cheap per unit because they are capital-intensive monopolies with decades of sunk cost to recover. That is why the meter never comes off, and why non-payment gets you disconnected.
The model-year trap is therefore not that AI fails to get cheaper. It is that executives may treat falling prices as a reason to delay, standardize, or lock in vendor commitments, when the strategic question is actually which capabilities are depreciating and which still justify frontier investment.
For most applications, where the needed strategic response is speed and cost discipline, the answer increasingly will be the free tier. Only a small number of applications will deserve the frontier budget â and the CEOâs attention.
Authors
Related
AI may write the recipe, but only we can taste the food
AI can optimize processes, but true leadership requires human judgment, says Shelley Zalis. Success depends on integrating machine intelligence with âsoft skillsâ like empathy and intention....
Audio available
AI and the CIO: From Chief Information Officer to Chief Intelligence Officer
Facing a fundamental redefinition of their purpose in organizations, todayâs CIOs need to pivot quickly. ...
How Swarovski used GenAI at scale to generate real profit
The Austrian crystal company achieved over 5% contribution to digital commerce with less than 1% of digital budget by emphasizing a value-first approach, its Chief Digital and Information Officer told IMDâs Luxury...
Learn Brain Circuits
Join us for daily exercises focusing on issues from team building to developing an actionable sustainability plan to personal development. Go on - they only take five minutes.
Â
Explore Leadership
What makes a great leader? Do you need charisma? How do you inspire your team? Our experts offer actionable insights through first-person narratives, behind-the-scenes interviews and The Help Desk.
Â
Join Membership
Log in here to join in the conversation with the I by IMD community. Your subscription grants you access to the quarterly magazine plus daily articles, videos, podcasts and learning exercises.
Â