Dodgy bets on the future

Large AI platform providers are scrambling to launch their IPOs as soon as possible, to catch what remains of global liquidity – and valuations are breathtaking. Excluding the smaller IPOs (which would be huge in any other context), OpenAI and Anthropic are aiming for about $1tn each while SpaceX, now branding itself an AI-first company, is going for $1.75tn.

The latter is a particularly ambitious target not just due to its size but since its AI model (which currently holds under 4% of the market) needs rebuilding “from the foundations up” and its Starship rocket is yet to demonstrate a fully reusable, reliable, operational launch system. According to its own IPO filing, the company’s only profitable segment is its satellite internet division. In Q1 2026, Starlink generated $1.19bn operating profit, while SpaceX as a whole made a $1.94bn operating loss. The AI division alone lost $2.47bn.

Of course, all these IPOs are bets on the future. But it’s worth bearing in mind that AI is unlike traditional software, in which the costs are largely upfront. Traditionally, one invests massive initial effort in building a software product then scales back to maintenance / upgrade while selling endless copies at low marginal cost. Even with cloud services, the product becomes cheaper to serve as more users join, because each additional customer mostly consumes the same code. With LLMs, the cost of token consumption and generation never gets less, because every query requires live computation, and the most valuable use cases often involve longer prompts, larger models, multi-step agent workflows, retrieval, tool calls, and repeated generation.

In fact, there’s an argument that margins may decrease as an LLM cloud service grows. Providers can gradually reduce the cost per token through better chips, batching, quantisation and caching, but the more successful a platform becomes, the more it must spend on inference capacity, energy, networking, memory and infrastructure. LLM providers can box clever by improving inference efficiency, routing work to smaller models, charging more for heavy usage, or shifting users toward architectures that treat tokens as a scarce resource, but this isn’t the current approach by any means. Right now, LLM cloud providers are covering over 90% of token costs themselves.

So what underpins the landmark valuations of AI model providers? I see four possible explanations: streamlining routine human effort, enabling complex human effort, making scientific advances, and a military arms race. Of these, only one really stands up.

1. Streamlining routine human effort

This is the focus of most pilot adoption projects: the much-publicised efforts to eliminate low-level knowledge work (commercial transactions, financial checks, legal review, software coding, etc.) by replacing people with semi-automated agentic workflows. Even with the massive discounts currently being provided by LLM cloud providers, ROI from real-world adoption projects so far is not encouraging.

PwC’s 29th Global CEO Survey found that more than half of CEOs had seen neither revenue growth nor cost reduction from AI, while an NBER survey of nearly 6,000 executives found that more than 80% of firms reported no impact on productivity or employment over the previous three years. MIT’s “GenAI Divide” report was even harsher, arguing that most enterprise GenAI pilots have produced no measurable P&L impact because tools fail to learn from context, fit workflows, or scale beyond individual productivity.

There are success cases. Klarna reported a 40% reduction in customer-service cost per transaction after deploying AI, and Walmart reported that an AI-directed workflow reduced shift-planning time from 90 minutes to 30. However, these organisations achieved results via highly disciplined implementation projects (workflow redesign, cost control, contextual learning, governance and integration into everyday operations) that may be beyond the capability of many smaller organisations.

Most organisations never really got to grip with enterprise business process management in the first place, and this now seems like a prerequisite for successfully replacing low-level staff with AI. If anything, it seems as if enterprise process automation may be abandoning both BPM and AI and moving towards a traditional approach to paving over cowpaths.

2. Enabling complex human effort

By contrast, there is a powerful opportunity for AI at the other end of the human effort scale: the tricky stuff, where subject matter experts do deep thinking to analyse complex challenges and respond appropriately. Areas requiring this kind of thoughtful work include safety assurance, risk management, crisis planning / response, strategy development, architecture (physical and digital), and so on.

This is a key focus for us with Dedoctive. Our auditable AI for defensible decisions wraps LLMs to produce sophisticated outputs that can be checked, challenged, and defended. We recognise that in many domains, there are not enough skilled people or enough time available to do this kind of work properly. There never has been – that’s why all across the globe we still end up with major projects that over-run, change programmes that under-deliver, engineering systems that fail, and (perhaps scariest of all) technology deployments with unquantified risks.

It’s not only a resource issue. For example, to properly assure a safety-critical system such as an aircraft, human decision-makers must somehow synthesize a vast amount of documentation, identify areas of concern, and work through the implications of usage in multiple operating environments – including in unexpected operating environments that were not included the original design parameters, since this can and will happen, like it or not. A tool such as the Dedoctive Safety Navigator relieves some of the cognitive strain, allowing humans to use their skills optimally.

This is important work, especially in a time when deployment of new technologies such as nanotech, biotech, geo-engineering, and AI itself could accidentally wipe out humanity. However, mainstream AI cannot at this time be trusted to assist in such important decisions. Big tech companies do not appear even to see remedying this deficiency as a priority. The closest they get is so-called “thinking” models that iterate repeatedly over a task with the aspiration of reducing inaccuracy and responses that append a set of reference links (some of which may be relevant).

This is partly due to the limitations of the neural net technology underpinning LLMs (which can only be remedied via a neuro-symbolic platform such as Dedoctive), and partly due to the fact that most managers simply don’t care enough about this type of work. Verification and validation (V&V) work is under-funded in most organisations since humans are prone to optimism bias. As Wikipedia puts it, the tendency “is common across cultures, genders, ethnicities, nationalities, and age groups”.

Almost all humans believe that cost-cutting on assurance makes sense because, hey, what are the chances of the worst happening? And we don’t seem to learn from experience (e.g., of major global Black Swan events such as financial crises and pandemics) that asking about the chances of such events is the wrong question. We should be asking about, and proactively mitigating, the impacts of such events.

So betting that this type of work will make AI profitable at the grandest scale is effectively gambling on human nature changing at a fundamental level. Well, it might pan out, but a million years of human evolution are lining up to take your bet, at any odds you care to give.

3. Making scientific advances

AI is already giving us scientific breakthroughs. Use of AI to predict 3D protein folding has transformed structural biology, drug discovery and molecular biology workflows. The AlphaFold database now contains predictions for more than 200 million protein 3D structures based on their amino-acid sequence, and in 2024 AlphaFold3 extended the original approach from individual proteins to complexes involving proteins, nucleic acids, small molecules, ions and modified residues.

AlphaFold won the 2024 Nobel Prize in Chemistry for Demis Hassabis and John Jumper. They shared it with David Baker for his work on computational design of new protein molecules. In related work, Google DeepMind’s GNoME has predicted around 380,000 new stable crystal structures and Berkeley Lab has used AI to accelerate the synthesis of inorganic materials. These and other AI-enabled advances in chemistry will bring new drugs and technologies to market rapidly over the coming years.

However, advance in other fields has been slower. For example, there were great expectations for AI in mathematics, which practitioners have long suspected was amenable to automation, but so far AI has not solved any of the great open research problems such as the Riemann Hypothesis, P vs NP, or Navier–Stokes. Humanity’s Last Exam (HLE) shows that the average post-graduate outperforms all Frontier AI models. My guess would be that the trend of future HLE score improvements is sublinear as the remaining questions become harder, extra inference compute has diminishing returns, and benchmark noise becomes more significant.

Returning to IPOs and future profitability, to what extent is scientific advance a priority for business spending? Ask any scientist how easy it is to obtain funding for foundational research, and the answer is likely to be a derisive laugh. Yes, businesses do R&D, but to obtain shareholder approval for the cost it is generally focused on innovation (development of potentially profitable new products) rather than on blue sky research.

Let’s turn to the last, and most promising, hope for future AI funding.

4. A military arms race

ENIAC, one of the first general-purpose electronic digital computers, was funded by the US Army Ordnance Corps during World War II to calculate ballistic firing tables. Britain’s Colossus went into operation at Bletchley Park in 1944 to break the German Enigma code. The Internet’s precursor, ARPANET, was funded by the US Department of Defense in 1969 to connect geographically separated, Pentagon-funded research computers during the Cold War.

On 9 January 2026, the Pentagon’s MEMORANDUM FOR SENIOR PENTAGON LEADERSHIP COMMANDERS OF THE COMBATANT COMMANDS DEFENSE AGENCY AND DOW FIELD ACTIVITY DIRECTORS stated:

In the national security domain, AI-enabled warfare and AI-enabled capability development will re-define the character of military affairs over the next decade. This transformation is a race – fueled by the accelerating pace of commercial AI innovation coming out of America’s private sector. The United States Military must build on its lead over our adversaries in integrating this technology … to make our Warfighters more lethal and efficient. To this end, aligned with America’s AI Action Plan, I direct the Department of War to accelerate America’s Military AI Dominance by becoming an “AI-first” warfighting force across all components, from front to back.

The US military is now implementing this via new and expanded contracts with some of the biggest names in technology. Under eight agreements with Google, OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia and the start-up Reflection, the Pentagon said AI technology would now be used for any “lawful operational use”. Given the current administration’s cavalier interpretation of the word “lawful”, this is essentially carte blanche to build and deploy more killer robot bees, birds, and dogs.

Have you seen the Black Mirror episode Metalhead? If you would like recurring nightmares about this stuff, or just to understand some of what front line soldiers in conflicts such as Ukraine are facing in the real world, check it out. Autonomous and almost unstoppable death machines are coming en masse, and it’s only a matter of time before non-state actors get hold of them too. America sees little choice but to up the stakes in this terrifying move to asymmetric warfare since China and Russia are so far ahead in deployment of AI to support conventional warfare.

In and around the South China Sea, China is building an integrated unmanned maritime sensing network – a system of systems including unmanned underwater vehicles, gliders, autonomous surface vessels, research motherships, fixed sensors, floating platforms and island/reef infrastructure. Russia is not so advanced, but is using new vessels including unmanned submersibles together with special-purpose seabed-warfare vessels and shadow fleet merchant ships around the Baltic to build persistent undersea awareness with the capability for grey-zone coercion activities such as exploiting cable / pipeline vulnerabilities.

Of course, the US is also building unmanned military platforms, but is far behind in their deployment. For example, delivery of the Orca XLUUV, an extra-large autonomous undersea vehicle, was delayed – it remains a prototype-heavy programme – and the Manta Ray UUV is still in testing. However, at this point in time the US has a structural advantage in compute concentration – if not in models, where US and Chinese models have been trading the lead since early 2025 (and as of March 2026 Anthropic’s top model was only 2.7% ahead of the leading Chinese model). The US compute lead is unlikely to last for long, and depends of course on what happens to Taiwan over the next few years, so the US military is doubling down on AI while it still has an advantage.

Conclusion

The big question is whether or not current API IPOs are a bubble waiting to burst. If you agree with my argument above, this may come down to whether or not the US can and will sustain military funding for AI compute resources.

The IMF expects US federal deficit to exceed 6 percent of GDP in the next few years, and the federal debt-GDP ratio to steadily increase to 140% by 2031. These are startling figures. The 1946 peak of debt-GDP was only only 106%. At that time, deficit was 7% but this became a surplus of 1.7% by 1947, and the idea that 2027 or any upcoming year will see a US government surplus is ludicrous.

To put debt-to-GDP on a downward path, the IMF says a “clear, frontloaded fiscal consolidation plan” is needed: wonk-speak for tighten your belt and do it now. However, the current administration clearly sees military funding for AI compute as an existential priority. So, the viability of current AI IPOs may come down to whether the US or China is more financially capable of dominating the world stage over the next few decades.

Chinese debt is officially much lower than the US, but once broader public-sector liabilities are included, the IMF estimates China’s augmented debt to be 8.2%. This is is 135.3% of GDP, projected by the IMF to rise to a staggering 153.7% by 2030. They link China’s debt-sustainability challenge not to AI but to off-budget investment, industrial-policy support, and Local Government Financing Vehicles. However, China’s RMB 2tn ($295bn) five-year plan for national AI data-centres can’t be helping.

AI may be leading the world’s greatest nations into a death spiral of unchecked spending on technology that improves productivity only for assurance work about which no-one really cares – even if they should.