ChatGPT in 2026: already the end of the promised hegemony?

As a digital marketing professional for more than 15 years, I have had a front-row seat on the breakneck evolution of artificial intelligence (AI). For two years the Silicon Valley has been selling us an almost messianic vision: a single, omnipotent artificial intelligence, able to do everything better than a human.

But what if that vision were a decoy? What if, as in medicine, real excellence lay in specialisation?

My conviction, backed by technical and economic analysis, is that we are watching not the birth of a monopoly but an inevitable fragmentation of the AI landscape. That obviously calls into question the current overvaluation of the “Magnificent Seven” (Apple, Microsoft, Alphabet, Amazon, Meta, Tesla and Nvidia), who are positioning themselves as the future major players in a sector that remains speculative, fuelled by circular investment and facing colossal energy challenges. Let us look at it together.

Updated 4 September 2026. I published this article on 2 January 2026 with a market scenario that stopped in March. Eight months on, I am updating it for three reasons. The fragmentation I predicted happened far faster than I expected. The catalysts I had dated came and went without causing a break, which deserves to be said just as plainly as the rest. And above all, the main risk factor has changed in nature.

The myth of the “universal doctor”: why a “generalist” AI cannot do everything

An AI that would be surgeon, radiologist, psychologist and general practitioner all at once — the digital “universal doctor”. The financial markets, fascinated, pushed the valuations of AI’s “Seven Wonders” (Nvidia, Microsoft-OpenAI, Google and the rest) to stratospheric levels, anticipating a future monopoly.

Think of AI as the medical world. A general practitioner is valuable for an initial diagnosis and an overall view, but for complex problems you turn to specialists: a cardiologist for the heart, a neurologist for the brain, an oncologist for cancer. In the same way, a generalist AI such as ChatGPT (OpenAI) excels at everyday tasks — basic writing, brainstorming — but struggles with sharper needs, where specialised AIs outperform it.

Midjourney : still a premium reference for artistic rendering and for an “aesthetic signature” much appreciated by creatives, with a Discord ecosystem that remains active. But its leadership is no longer beyond question: the image models built into Gemini and ChatGPT have largely closed the gap in the independent 2026 rankings.

Claude (Anthropic) : the specialist in programming and long-form writing. It stands out for its precision on complex code (debugging advanced algorithms, for instance) and its ability to structure in-depth content, with fewer hallucinations. It is the clearest riser of the year: from 2% market share at the end of 2025 to around 10% by the summer of 2026, that is close to 245 million monthly users, driven by professional use.

Perplexity : AI’s “academic researcher”, expert at reliable sourcing and citations. It is also the most useful counter-demonstration of 2026: its share fell from 5.3% in January to around 2% in August, simply because ChatGPT, Gemini and Claude built sourced search in natively. Specialisation protects you only as long as a generalist cannot absorb it. That is the nuance I did not state firmly enough in January, and it is worth correcting.

Grok (xAI) : the specialist in instant information and social trend analysis through X (formerly Twitter). Perfect for insight on real-time events. Its share has held steady at around 2.8% in 2026, in a market that has otherwise moved a great deal — its niche protects it as much as it caps it.

Gemini (Google) and AI Mode: the big winner of 2026. A champion of multimodality, but above all of distribution: Search, Workspace, Android, Gmail. Its share went from 18.2% at the end of 2025 to 27.6% in September 2026, with the strongest quarterly growth in the sector (+17%). It is not winning by being better, it is winning by already being there — which, for anyone who works in search, is no surprise at all.

And why not a super generalist AI that would absorb the others?

Optimising for everything dilutes excellence. Just as a doctor cannot master every speciality without years of study, scaling an AI to excel everywhere demands massive resources in data and compute. Niches (secure code, sourced research) need dedicated architectures.

The figures confirm it, and faster than I was writing in January: ChatGPT went from 87% of the market in 2024 to 68-74% at the end of 2025, then to 51.5% in September 2026 according to First Page Sage. Another measure, Sensor Tower’s, centred on mobile usage, already put it below the 50% mark in May 2026, at 46.4%. The methodologies differ; the direction does not.

One point deserves to be made clear, because it is consistently misread: in absolute terms ChatGPT is still growing, with roughly 1.1 billion monthly users against 662 million for Gemini and 245 million for Claude. It is not shrinking, it is growing more slowly than the market. That is the exact signature of a loss of hegemony, and it is far harder to correct than a drop in visitors.

An alternative, and a new challenge: “the meta-engine that aggregates the AIs”

Rather than switching between 10 AIs, imagine a “new Google”: a search engine that queries several AIs in parallel, assesses their credibility, compares the answers to surface alternatives, and ranks the whole thing intelligently.

The results page would mix organic search (organic links) and AI blocks: a title, a short description, and a link to the full answer. That would solve fragmentation without a monopoly, and democratise AI without depending on a single giant. As things stand, though, my vision of an AI meta-engine runs into three major locks.

Technically, the cumulative latency and the synthesis of heterogeneous answers make the experience slow and confusing.

Economically, the cost per query, multiplied across several models, is prohibitive at scale, with no viable business model. Could the arrival of high-performing open-source models change that by making aggregation feasible?

Strategically, will the AI giants ever accept being content suppliers inside a meta-engine? And would a meta-engine not itself become a more powerful monopoly than the ones it claims to bypass? Only time will tell.

A speculative bubble, circular investment and overvaluation of the Seven Wonders?

The Magnificent Seven are selling themselves as the future giants of AI. Take an example you all know: Nvidia, the leader in the GPU chips that AI depends on. Its share price exploded on the back of circular deals with OpenAI. Nvidia sells billions in chips to OpenAI, which attracts massive investment, which in turn inflates demand for Nvidia chips.

Eight months on, it is the gap between the results and the share price that has become the most interesting piece of this file. In May 2026 Nvidia published record quarterly revenue of 81.6 billion dollars, up 85% year on year — and the stock fell the next day. Since its peak on 14 May 2026 the shares have lost around 18%. Three signals explain the gap: the collapse in GPU rental rates, with the B200 going from around 6.11 to 4.22 dollars an hour between late May and late June 2026, a 31% fall in three weeks; data centre revenue in China down to virtually zero under export controls; and sustained insider selling over the quarter.

My reading is this: the market has stopped paying for growth and started paying for the durability of that growth. It is not a crash, it is multiple compression. Historically that is what precedes one — without it being a rule.

On the OpenAI side, the valuation went from 500 billion dollars in October 2025 to 852 billion in March 2026, on an annualised revenue run rate of around 24 billion. The IPO, once floated for 2026, has been pushed to 2027 — officially to aim at the 1,000 billion mark, in practice because the market now demands answers on real profitability. A valuation-to-revenue ratio of around 35 poses a simple question: what is being bought, a company or a hypothesis?

It is a “circular economy” described by analysts: the same capital circulating between giants, inflating bubbles with no concrete near-term profit. Michael Burry (famous for predicting the 2008 crisis) accuses Nvidia of feeding that bubble, with misdirected investment.

That diagnosis is no longer the preserve of short sellers. In August 2026 the European Central Bank published a post arguing that a correction in US technology valuations is likely, even on the assumption that AI delivers every one of its technical promises. It puts euro-area households’ direct exposure to the Magnificent Seven at around 440 billion euros, with a comparable amount for the insurance and pensions sector. Its most uncomfortable argument is not about the level of valuations but about the capacity to absorb a shock: unlike in 2000, central banks today have far less room to cut rates or support the economy.

The irony of the subject: Nvidia established itself precisely through extreme specialisation. In other words, even Nvidia illustrates the “specialist” logic. It did not win by being a generalist, but by becoming the best in one precise domain.

The gap between the sector’s revenue and the amounts committed to infrastructure has narrowed over the past year, but it remains an order of magnitude: spending on compute centres continues to grow faster than the revenue it generates, and the break-even point of the 2024-2026 investment wave is not expected before the end of the decade. It is precisely that time lag that makes the sector sensitive to the cost of capital — and so to the Fed’s decisions.

Unlike the dot-com bubble (built on web promises with no infrastructure), the AI bubble is held back by tangible constraints such as operating costs and energy.

The energy challenge: the sinew of a war that has already begun

Imagine that all the data centres on the planet today consume as much electricity as France and Spain combined. That is the reality of 2026, and the slope has steepened again since January: consumption rose 26% in twelve months. By 2030 it could more than double.

To grasp the scale: every time you ask ChatGPT a question it uses 2 to 6 times more energy than a simple Google search. Multiply that by billions of daily queries and you see the problem.

The real difficulty? The infrastructure is not keeping up. In the United States, experts forecast a shortfall in electrical capacity equivalent to 35-50 nuclear power stations by 2030. In Europe, building a data centre takes 2 to 3 years between permits, environmental studies and grid connection. An eternity when the technology moves in months.

For the more technical among you: in June 2026 Gartner put data centre consumption at 565 TWh for the current year, against 447 TWh in 2025, for a power draw of 133 GW against 105 GW a year earlier. The United States accounts for 204 TWh of that, including 68 TWh for AI-dedicated data centres alone. By 2030 the projection passes 1,200 TWh and 291 GW, with AI servers making up close to half the total. The 945 TWh I cited in January were therefore a low estimate.

In short, the exponential growth AI promises runs into an elementary equation: available electricity. Without a revolution in the way we produce and distribute energy, the promise of an infinitely scalable AI risks meeting a very real wall.

China builds reactors in months, the West in years: what that means for the AI bubble

Some countries are faring better. China, with its fast growth in solar and nuclear (including thorium, cheaper, less radioactive, and producing less waste than uranium), holds 25% of the world’s data centres against 45% for the United States. Less hampered by regulation, it builds power stations in months rather than years, securing its AI lead. In November 2025 it carried out the first thorium-to-uranium conversion in a molten salt reactor, with a target of 150 reactors by 2050.

Europe and the United States risk falling behind, and a sequence of events could be enough to test the “infinite AI growth” narrative — with, potentially, a brutal repricing of the most exposed stocks. Let me state the frame deliberately: I am not a trader and what follows is neither a prediction nor investment advice. It is an illustrative scenario, built on dated public catalysts that historically tend to cause market adjustments. My point is not “this will happen” but “here is how it could happen” if several signals stack up.

What my January scenario got wrong

In January I placed this sequence in February-March 2026: Nvidia’s results on 25 February, GTC in mid-March, a Fed meeting. None of the three triggered the break I described. Let me say it plainly: I got the calendar wrong, and that is worth writing down — a scenario that does not play out on the expected date does not become false, it becomes better informed.

What these eight months have shown is that the adjustment did not come through an event, but through erosion: Nvidia’s stock has lost 18% since May despite record results, which is a far slower and far less legible way to correct than a drop on a publication. And above all they have shifted the nature of the risk. In January, the limiting factor I identified was electricity. Today it is the geography of production, with the tensions around Taiwan in particular. So here is the calendar as I see it for the end of 2026.

The scenario, readjusted: fourth quarter 2026

• 27-28 October 2026: Fed meeting. No economic projections, so limited in itself, but it sets the tone before the November sequence. A more restrictive message raises the cost of capital, and tech valuations rest precisely on distant cash flows, so they are very sensitive to the discount rate.
• 17 November 2026: Nvidia’s fiscal Q3 2027 results (date confirmed, after close). The central appointment. The market will not be looking at revenue, which will probably be another record, but at two lines: whether GPU rental prices stabilise or not, and the visibility given on the rollout of the Vera Rubin platform. Cautious guidance on those two points would strike at the heart of the “AI trade”.
• 28 November 2026: local elections in Taiwan (the so-called “9-in-1” vote). Eleven days after Nvidia. These local elections are no longer local: Taiwan’s National Security Bureau recorded a 60% rise in inauthentic social media accounts between 2024 and 2025, and more than two million instances of disinformation in 2025 alone. A sequence of military or informational pressure around that date would put Taiwan risk back at the centre of the table — and with it the entire AI chip supply chain.
• 8-9 December 2026: Fed meeting with projections (the dot plot). The last of the quarter, and the only one to publish rate projections. In a market already jittery after November, that is the catalyst capable of turning a sector correction into a broader repricing.

Taken individually, each of these events is absorbable. What interests me is their concentration into six weeks and, above all, the fact that they are no longer independent of one another: a Taiwanese incident mechanically degrades Nvidia’s visibility, and a constrained Fed reduces the system’s capacity to absorb shocks. That is the definition of rising correlation — the moment when diversification stops protecting you.

And I will say it again, because the nuance matters: I am describing a possible chain of events, not a forecast. The most likely scenario is still probably the one of the last eight months — a slow erosion, with no identifiable break.

Today AI is no longer an isolated “tech” subject: it is a narrative pillar of the markets. And because I work with AI daily, I also see its constraints (costs, energy, infrastructure, maturity of use). Things that can, at certain moments, catch up with the financial storytelling.

The real breaking point is no longer energy, it is the Taiwan Strait

In January I concluded that AI’s wall would be electrical. I stand by the diagnosis, but I rank it differently. An electricity shortfall is a problem of pace: it slows things down. A supply chain break is a problem of existence: it stops them.

One figure sums the situation up: TSMC concentrates in Taiwan around 95% of the world’s advanced CoWoS 2.5D packaging capacity — the assembly step without which an AI accelerator simply does not exist. By the end of 2026 the foundry should reach 120,000 to 140,000 wafers a month on that process, against around 70,000 in 2025, and as early as late 2025 Nvidia had reserved between 50 and 60% of that 2026 capacity. TSMC’s US plants will not relieve the bottleneck before 2028-2029.

In other words: almost all the world’s AI compute for the next three years depends on an industrial process concentrated on an island roughly 180 kilometres off the Chinese coast. That is not a geopolitical opinion, it is a production fact.

Where do things actually stand in autumn 2026?

The summer sent contradictory signals, which is exactly what makes it hard to read. On the pressure side: a record 244 Chinese vessels recorded in Taiwanese waters in July. On the apparent easing side: 125 incursions into the air defence identification zone in August, a notably lower level. From Taipei: a supplementary budget of 240 billion Taiwan dollars voted for drones, and a referendum approved on 28 August opening the way to a return to nuclear power, which would reduce dependence on imported liquefied natural gas — and so the leverage an energy blockade would offer. From Beijing: fresh purges in the high command, with two generals removed.

My reading, as an observer and not as a geopolitical analyst: the front line has moved from the military to the cognitive. The most likely scenario for late 2026 is not an invasion, it is a large-scale influence campaign around the 28 November vote, supported by AI systems built for the purpose. That is no more reassuring: markets do not price an invasion, but they price uncertainty very badly. A simple sequence of tension would be enough to lift CoWoS concentration risk out of the footnote where it has been filed for two years.

There is an irony here I find hard to ignore: the most strategic infrastructure of this century rests on a single point of passage, and AI is simultaneously the tool being used to weigh on the election of the territory that houses it.

And what if Elon Musk were once again a step ahead…

A bold vision: data centres in orbit through Starlink V3, announced in October-November 2025.

The advantages: unlimited solar energy (closer to the Sun), complete autonomy, and zero terrestrial regulation. No waiting for permits, no energy shortages. Technically it rests on high-speed laser links for communication (latency ~20-30 ms against terrestrial).

But major challenges: launch costs (billions, via Starship), heat dissipation in the vacuum of space, and robotic maintenance.

In the medium term (2030+) specialists consider it credible. Musk sees it as an extension of Starlink for AI payloads. Bezos is following with Blue Origin, making it an AI space race.

Conclusion

After months of testing, analysing and observing AI day to day — in my work, my personal projects, even in conversation with people around me — I have to admit to a small disappointment. We were sold a dazzling “new era”, an upheaval comparable to humanity’s great ruptures. Yet when I look back, the real revolutions touched everyone, deeply.

The discovery of fire let humanity survive, cook food, keep warm; every individual benefited. The invention of the wheel and of stone-age tools transformed work and travel for all. The industrial revolution, with the steam engine, the railways and electricity, connected continents, created mass employment, urbanised societies and changed the daily lives of billions. More recently, the arrival of the internet, Google and smartphones democratised information, instant communication and services. My grandmother uses WhatsApp, my father looks everything up on Google, and whatever one thinks of it, children navigate tablets from a very young age.

AI, for now, does not have that universal impact. It remains above all a powerful tool for professionals, creatives and developers — people who already have a precise objective. In my own circle, as in the recent statistics, the general public tries it out of curiosity but has not yet made it part of daily life as an indispensable reflex.

That does not mean AI is useless — quite the opposite. It is already hugely beneficial and will be more so: more precise medical diagnosis to treat cancers and rare diseases, safer roads and cities, faster discovery of new drugs, help with the energy transition through predictive models. It will save lives, reduce inequality in access to knowledge and boost productivity.

My conviction is therefore this: AI is a major, beneficial, structuring advance… but it does not yet have the universal, immediate character of the great mass-market revolutions. Perhaps the real “AI revolution” will come when it can no longer be seen: built in everywhere, effortlessly, the way web search has become. In the meantime, let us keep our clarity: we are at the start of a long trajectory, and that is precisely why it needs to be thought about, governed and used intelligently.

What eight months have changed in my analysis

In January I thought the AI market would fragment: it did, faster than I wrote. I thought specialisation would protect the niche players: Perplexity demonstrated the opposite, because a speciality a generalist can absorb is not a defensible position. I thought the main constraint would be energy: it is first industrial and geographic, and it sits in an assembly process concentrated on a single island. And I thought a correction would come through quarterly results: so far it has come through a slow erosion, with no trigger event.

I will revisit this at the end of the year, once the November-December sequence has passed. That is the principle of an article you keep up to date rather than one you archive.

Sources used:

• Jedha — comparison / market share (December 2025)
• Unite.AI — market share and growth summary (December 2025)
• Similarweb — traffic and share data (cited via Unite.AI)
• The Decoder — market share / Grok / trends (December 2025)
• FirstPageSage — market share / traffic (December 2025)
• Exploding Topics — growth and traffic indicators (December 2025)
• Grace Blakeley — Substack (October 2025)
• Yahoo Finance — analysis / Michael Burry / Nvidia (December 2025)
• The Atlantic — analysis (December 2025)
• Sequoia Capital — revenue against infrastructure investment estimates (2025)
• International Energy Agency (IEA) — data centre consumption / 2030 projections (2025)
• Goldman Sachs — electrical capacity shortfall / projections (2025)
• World Nuclear News — thorium / molten salt reactors (November 2025)
• NucNet — thorium / nuclear (November 2025)
• PowerMag — nuclear / projections (December 2025)
• Forbes — data centres in orbit / technical constraints (December 2025)
• SingularityHub — space data centres / limits (December 2025)
• NY Post — the AI space race (December 2025)
• Space.com — space race / feasibility (December 2025)

Sources added for the September 2026 update

• First Page Sage — AI chatbot market share (September 2026)
• Sensor Tower — State of AI Report 2026, mobile market share (May 2026, cited via TechCrunch)
• TIKR — Nvidia analysis: share performance, GPU rental rates, China exposure (June 2026)
• Wall Street Horizon — Nvidia earnings calendar (2026)
• US Federal Reserve — 2026 FOMC meeting calendar
• European Central Bank — post on the expected correction in tech valuations (August 2026)
• Gartner — data centre electricity consumption, 565 TWh and 133 GW (June 2026)
• American Enterprise Institute — China & Taiwan Update (1 September 2026)
• Global Taiwan Institute — Taiwanese local elections and cognitive warfare (May 2026)
• Taipei Times — Taiwan’s vote set for 28 November 2026
• Sector analysis of TSMC’s advanced CoWoS / SoIC packaging (August 2026)
• French business press — OpenAI’s valuation and IPO delay (2026)

Written by

Nicolas Peter – Performance Web and SEO expert in Perpignan and Paris
Oli-via-net agency