Don’t put your future in one AI basket: Nokia Risk, the real economics of ChatGPT, and why prudence is strategic

Apple's first-ever mobile phone was the iPhone

The corporate rush into large language models is not driven by curiosity. It is driven by pressure. Boards are asking what the AI plan is. Clients are asking whether delivery will be faster. Employees are asking whether roles will be redesigned around new tools. In several sectors, AI has become a relevance test, not an innovation option.

That is exactly why the most valuable posture right now is not blind acceleration, but disciplined prudence.

Two bets that leaders keep conflating

Most organisations are making a subtle mistake. They conflate two different bets and treat them as one.

The first bet is the ecosystem bet: that AI will be a durable part of how knowledge work gets done.

The second bet is the dependency bet: that a specific vendor, model family, contract structure, data pathway, and operating model is safe enough to anchor critical work on for years.

The ecosystem bet is increasingly hard to argue against. The dependency bet is where the unpriced risk sits.

Nokia Risk: when the category rules change under your feet

This is where a phrase helps leaders keep their judgement when enthusiasm starts running ahead of governance: Nokia Risk.

Nokia Risk is what happens when an organisation commits deeply to the wrong assumptions about a fast-moving category, builds capability around those assumptions, and then discovers the category rules have shifted. The painful part is that the organisation may have executed well by its internal measures. It simply executed well inside a frame that stopped being the right frame.

The Nokia analogy is not about whether anyone “deserved” the outcome. It is about discontinuity.

Apple unveiled the iPhone in January 2007 and launched it in June 2007, catalysing a shift from phones as hardware products to phones as software platforms and ecosystems. Nokia’s handset position did not vanish in a week, but the category logic changed quickly enough that previous strengths became less decisive. Nokia ultimately sold its Devices & Services business to Microsoft in a deal announced in September 2013.

The strategic lesson is not “never back the incumbent”. The lesson is that when the rules change, accumulated advantage can turn into inertia, and inertia can become fragility.

AI is creating the conditions for a similar discontinuity, but with an added twist: the dependence is not only technical. It is behavioural.

It shapes how people ask questions, how they write, how they verify, how they collaborate, and what they treat as “good enough”. This is not like being locked into Microsoft 365 for documents and spreadsheets, where lock-in is mostly a tooling and file-format problem. LLM lock-in is an operating model problem. It reaches into the way decisions are formed and the way work is coordinated.

That is why “don’t put your future in one AI basket” is not an anti-AI slogan. It is a governance slogan.

Prudence is not hesitation. It is separation.

The practical posture is simple: separate capability development from dependency. Build the skill, the controls, and the workflows, but avoid making a single vendor pathway the spine of your business.

The lesson is not “never back the incumbent”. The lesson is that when the rules change, accumulated advantage can turn into inertia, and inertia can become fragility.

The hard maths: what ChatGPT’s economics imply for buyers

Prudence is easier when you look directly at the incentives of the ecosystem you are buying into. Those incentives are heavily shaped by capital intensity, infrastructure constraints, and a business model that is still being stress-tested in public.

OpenAI’s funding history is a useful lens because it makes the scale explicit.

In March 2025, Reuters reported that OpenAI would raise up to $40 billion in a SoftBank-led round at a $300 billion valuation, with Microsoft and other investors participating.

By February 2026, Reuters reported that Nvidia was nearing a $30 billion investment as part of a broader fundraise targeting over $100 billion, at valuations being discussed around the $830 billion range, and that OpenAI expected compute spend of around $600 billion through 2030. The Financial Times reported the same direction of travel on the Nvidia investment reshaping prior commitments.

Those numbers matter for a buyer because they point to a simple reality: the frontier LLM business is not “software with high margins” in the traditional sense. It is infrastructure-heavy, capital-hungry, and constrained by physical supply chains. Even if a vendor is brilliant, the physics still applies.

Now to the daily burn question, with the necessary honesty: there is no single, universally agreed “ChatGPT burns X per day” number that is both precise and publicly audited. But we can do grounded maths using reported projections and show what it implies.

Reuters reported in September 2025 that OpenAI expected to burn over $8 billion in 2025, with projections rising sharply in subsequent years (for example, $17 billion in 2026, $35 billion in 2027, and $45 billion in 2028), citing The Information.

Taking only the 2025 figure, a burn of $8,000,000,000 over 365 days is about $21,917,808 per day. That is not a moral judgement. It is a scale indicator: the cost structure is measured in tens of millions per day at the company level, before you even talk about the broader ecosystem of hyperscalers and chip suppliers that must expand to make the whole thing possible.

You might reasonably ask: is that “ChatGPT” or “OpenAI”? It is OpenAI’s business burn projection, which includes more than the consumer chat product. But it remains a useful public anchor because the same cost driver dominates: compute, especially inference at scale.

Reuters further reported in February 2026 that OpenAI’s inference expenses quadrupled in 2025 and that margins compressed as operating costs rose.

This is where prudence becomes practical for enterprises. When a supplier’s economics are that capital-intensive and still evolving, you should assume pricing, packaging, and contract structures will continue to change. Some of those changes may be favourable. Some will be driven by the supplier’s need to defend margins. That is not sinister. It is the reality of a rapidly scaling, physically constrained industry.

Energy and “per query” cost: the wrong question, but a useful signal

Sam Altman has publicly pushed back against inflated claims and has cited an energy figure per typical query on the order of a fraction of a watt-hour. Business Insider reported him citing about 0.34 watt-hours per query.

Whether or not you accept that exact number, the direction of travel is not in dispute: inference at mass adoption is where the energy and infrastructure load concentrates, not training runs once in a while.

The International Energy Agency projects that global data centre electricity consumption is set to more than double by 2030 to around 945 TWh, with AI a major driver.

The hard maths takeaway is not “AI is too expensive so avoid it”. The takeaway is that the industry is building a new layer of industrial infrastructure under what many people still talk about as if it were a normal SaaS product. That difference will show up in availability, pricing, regulation, and the vendor landscape.

The physical bottlenecks: why this is not just a software story

A prudent strategy respects bottlenecks, because bottlenecks are where surprises are born.

Bottleneck 1: power and grid connection. The data centre boom is colliding with real constraints: transformers, grid equipment, utility timelines, and community resistance. Axios reported that a significant share of global data centre projects slated to come online in 2026 may face delays, citing power constraints and grid equipment shortages. The Washington Post described how some firms are exploring off-grid generation strategies, including gas-powered setups, to bypass grid delays, introducing its own environmental and regulatory risks. In the UK context, Ofgem warned that proposed data centre projects could require power capacity on a scale that would materially strain the system.

Bottleneck 2: advanced chips and, increasingly, memory. GPUs are the headline, but high-bandwidth memory has become a gating factor for the newest accelerator designs. Fortune reported on how AI demand is intensifying pressure on the memory supply chain and that expanding capacity takes years.

Bottleneck 3: building the facilities. AI-optimised facilities require specialised cooling, power delivery, and high-density design. When you combine this with power constraints, it becomes clear why “we will just scale” is not a strategy. It is a hope.

These bottlenecks are not a sideshow. They shape your vendor risk. They also shape your architectural risk. If a provider is capacity constrained, you will feel it in prioritisation, latency, quota, price, and contract terms. If regulation tightens around energy, emissions, or data, you will feel it in compliance overhead and in what can be deployed where.

This is why Nokia Risk belongs in the AI conversation. Discontinuity often arrives through the constraints people ignore while they are excited about capability.

Why “AI productivity” is so hard to measure cleanly

The question many leaders are now asking is the right one: where is AI showing up in productivity data?

There is evidence of productivity gains in specific settings. But most organisations are not running controlled studies. They are running adoption programmes. Adoption programmes produce enthusiasm and anecdotes. Controlled measurement produces knowledge.

When consultancies and large firms talk about productivity gains, the crucial question is what they mean by productivity. Is it time saved on a task? Is it throughput? Is it quality-adjusted output? Is it fewer errors? Is it fewer escalations? Is it reduced rework?

If the measure is “time saved”, where did the time go? Into more volume, better quality, or simply more meetings?

The most common trap is to measure tool usage because it is easy. But usage is not value. Usage can be compliance theatre.

That is why Accenture’s reported approach of monitoring AI tool use and linking it to promotion decisions is such an interesting signal. It may well accelerate capability building. But if the metric is activity rather than outcomes, it can also produce a distorted sense of success.

A more prudent interpretation is to treat AI fluency as a leadership capability, but to demand proof through outcome metrics executives actually care about: cycle time, rework, quality, client outcomes, and risk events. This is not slower. It is more serious.

Why holding back can be the most strategic move

“Hold off” does not mean “do nothing”. It means you separate capability development from dependency.

You can invest heavily in AI literacy, governance, experimentation, and workflow redesign without putting your core operating model into one vendor’s pathway. That is the middle road between FOMO and denial, and it is where most serious organisations will end up.

It also allows you to market your expertise without lying to yourself. You can credibly say you are AI-enabled, AI-literate, and actively deploying use cases, while still refusing to confuse marketing posture with architectural commitment.

This is the distinction many leadership teams need: it is possible to be ahead in learning and disciplined in dependence at the same time.

A long-run operating model view

If AI is going to keep moving, your operating model needs to move with it. That is the core.

The strategic objective is not to “win AI”. It is to stay adaptable while others overcommit, and to avoid becoming the Nokia of AI adoption: heavily invested in the wrong assumptions, with no graceful exit.

If Nokia Risk is a useful phrase, it is because it forces a question that should be normal in every executive discussion of LLMs: If the category shifts in 18 months, how do we exit gracefully?

That single question changes what you build. It changes how you negotiate. It changes what you measure. It changes how you train people. It changes the role AI plays in your target operating model and your culture.

That is the prudent approach: not slower adoption, but smarter commitment.

References

1.       Altman, S. (reported 2026, February 24). Sam Altman says concerns of ChatGPT’s energy use are overblown; cites per-query energy use estimate. Business Insider.

2.       International Energy Agency. (2025). Energy demand from AI. IEA.

3.       Keynes, S. (reported 2026, February). Nvidia and OpenAI revise megadeal; funding round and infrastructure ambitions. Financial Times.

4.       Reuters. (2025, March 31). OpenAI to raise $40 billion in SoftBank-led round at $300 billion valuation. Reuters.

5.       Reuters. (2025, September 6). OpenAI expects to burn $115 billion through 2029, report says; 2025 burn and future projections. Reuters.

6.       Reuters. (2026, February 20). OpenAI expects compute spend of around $600 billion through 2030; inference costs rose; margins pressured. Reuters.

7.       Reuters. (2026, February 23). Bridgewater says Big Tech to invest about $650 billion in AI in 2026; infrastructure wave and risks. Reuters.

8.       The Washington Post. (2013, September 3). Microsoft buys Nokia device business for $7.2 billion. The Washington Post.

9.       The Washington Post. (2026, February 19). Silicon Valley building off-grid power for data centres; grid constraints and trade-offs. The Washington Post.

10.   Axios. (2026, February 23). Global AI data centre boom hits delays; power and grid equipment constraints. Axios.

11.   Fortune. (2026, February 15). AI demand drives memory constraints; HBM supply expansion takes years. Fortune.

12.   The Guardian. (2026, February 23). Ofgem warns data centres could materially strain UK power capacity. The Guardian.

13.   The Guardian. (2026, February 19). Accenture links promotions to use of AI tools, report says. The Guardian.

14.   Apple Newsroom. (2007, January 9). Apple reinvents the phone with iPhone (availability in June 2007). Apple.