July 23, 2026·13 min read13 min·

Massimo Mentrasti, Founder & CEO

Founder & CEO

Massimo Mentrasti

With three decades of experience working for international corporations, of which last two playing managerial and entrepreneurial roles, he has gained an unparalleled experience selling and delivering large transformation program across different industries and functional areas (e-commerce for retail, digital marketing for automotive, CRM and applications for banking, ground control system for aerospace industry). As Accenture partner during last 15 years, he has built from scratch and led the growth of practice focused on highly innovative offering as digital transformation, analytics, digital marketing and automation.

massimo@aevoluta.com

Two leading artificial intelligence companies, on the verge of going public, are together worth nearly two trillion dollars. The paradox is that they sell something that becomes, quarter after quarter, cheaper and easier to replicate. And as this happens, governments are starting to treat AI as a strategic weapon, to be rationed and granted.

They are OpenAI and Anthropic, and they are approaching the trillion-dollar valuation threshold.¹ It is curious that, exactly at this stage, the apocalyptic narrative about AI destroying jobs is softening: in May 2026 Sam Altman said he had been «quite wrong» about AI's economic impact, and Dario Amodei moved from «destroying» 50% of white-collar jobs to a more reassuring message about «multiplying output».² One possible reading is that the dystopian scenario, useful for raising private capital, becomes uncomfortable once a different audience needs convincing: institutional investors, regulators, and enterprise clients.

The rest of the picture completes the sketch. Google does not break out Gemini's revenue, buried inside search and cloud, but in April 2026 it committed decisively to the open field too, releasing the Gemma 4 family under an Apache 2.0 license — the first truly free one for commercial use.⁴ Meta raised its 2026 capex to $115–135 billion and made a symbolic pivot, moving from the open Llama to the proprietary Muse Spark model.⁵ DeepSeek surpasses $1.1 billion in revenue with prices close to cost; Mistral, the European alternative, is aiming for a billion after a round that valued it at $20 billion.

One thing the revenue figures hide is that neither of the two leading companies is profitable. OpenAI, according to The Information, burned through $3.7 billion in cash against $5.7 billion in revenue in the first quarter of 2026.⁶ The real profits sit elsewhere: upstream, with whoever sells the infrastructure (Nvidia, with gross margins around 70%), and downstream, with whoever already has a money machine to lean AI on (Google, Meta). And it is worth noting the irony: Meta, initially a sponsor of open models, is closing itself off around a proprietary model to chase the frontier just as the market discovers that open-weight is now good enough — while Google makes the opposite move, opening up Gemma 4 under a free license: the contest between open and closed does not follow a single direction.

A closed model works like a streaming service. Companies like OpenAI, Anthropic, or Google train the model, keep it on their own servers, and sell you access: you send a question through an interface or an API and get an answer back, exactly like pressing play and the film starts, without that film ever really being yours. You pay per use, usually in «tokens» (the units in which the model measures text — roughly three quarters of a word). The advantage is convenience: no server to manage, always the latest version available. The downside is that you depend entirely on the provider, for the price, for the terms of use, for service continuity, and for where your data ends up. And as with a streaming platform, the catalogue can change, the subscription can get pricier, and access can, in theory, be taken away.

An open model, or open-weight model, is instead like downloading the file and keeping it on your own disk. The model's «weights» — the billions of numbers that make up, in effect, its brain, the result of training — are published, and anyone can download them, run them on their own computers, inspect them, and adapt them. This is the path taken by models like Mistral's, DeepSeek's, or, historically, Meta's Llama. Here you manage the «disk space» yourself — meaning the hardware and expertise to run it — but the model is yours: once downloaded, no one can raise the price on you overnight, change the terms, read what you feed into it, or switch it off remotely.

Two clarifications worth keeping in mind. First: «open-weight» does not automatically mean «open source» in the full sense — in most cases, the finished model (the weights) is released, but not the training data or all the code needed to rebuild it from scratch: it's like having the film file, but not the original footage or the editing software to remake it from the ground up. Second: «open» does not mean «free». Downloading the model costs nothing, but running it well, at meaningful scale, requires machines and people, and those cost money.

Keeping this distinction in mind is what makes everything else readable. When the MIT study, further on, says that open models offer about 90% of the performance at one-seventh of the price, it is saying that the «file you keep at home» is now almost as good as the «subscription streaming» — and that is exactly the mechanism that turns language models into a commodity.

According to recent MIT research — a draft by Frank Nagle and Daniel Yue — across several benchmarks open-weight models reach around 90% of the performance of the best closed models, at a much lower inference cost: about 87% less in their sample ($0.23 versus $1.86 per million tokens).⁷ These numbers depend on the benchmarks and deployment conditions, not a universal law, but the direction is clear. The study adds an uncomfortable detail: in an economic simulation, not an observed measurement, shifting demand toward the dominant open models would save users an estimated $25 billion on a 2025 basis. The preference for closed models therefore looks more like organisational inertia than a stable technical advantage — though it isn't only inertia: reliability, support, contractual guarantees, and managed security remain concrete reasons why many companies continue to choose them.

And this is no longer just a market dynamic: it is becoming a declared position.

Coming from the CEO of OpenAI's main backer, while Microsoft itself is pushing low-cost, interchangeable models, this is a signal that carries weight: commoditization is no longer just something the market is doing, but something even the dominant players are starting to claim as necessary.⁸

Taken together, the three moves tell the same story: Google opens up Gemma 4, Microsoft openly calls for commoditization, Meta closes itself off against the trend. When giants move like this, commoditization stops being a prediction and becomes a direction.

A caveat, though: don't generalize. It is not language models as a whole that are becoming a commodity, but many of their most widespread, repetitive uses. Frontier reasoning, agent orchestration, multimodality, and deep enterprise integration can keep high margins for a long time yet. The market, rather than flattening, is bifurcating: a low-cost routine layer and a still-defensible frontier layer.

The clearest signal came on 12 June 2026, when the United States restricted foreign use of its own frontier models — in the most discussed case, Fable 5 and Mythos 5: an order justified on security grounds suspended access for any foreign national, forcing Anthropic to disable them for everyone.⁹ For years the implicit assumption was that artificial intelligence was a global service like any other: you pay and you access it, from any country. That block breaks the assumption. A leading model is no longer a product that anyone, anywhere, can simply buy; it becomes a strategic asset that a government can decide to withhold, ration, or grant, as is already done with «dual-use» technology, civilian and military at once.

The consequence runs in the same direction as the two forces described above. Cutting off access to American models does not slow down competitors: it forces them to build an alternative. It is the most powerful push imaginable toward the technological sovereignty of Europe and China. That sovereignty can take two forms: a national but still closed model, or an open model. And it is the second path that comes out strengthened, because an open model already downloaded onto your own servers is by definition much harder to «switch off remotely» than an API that depends on a foreign supplier. When a supplier's tap can be shut off by political decision, a model you already own becomes not only cheaper, but more secure in a strategic sense.

Two very different films are thus taking shape. On one side, some American companies on the verge of going public — Anthropic and OpenAI foremost — seem to be racing to build «the model that will dominate the world»: a single frontier intelligence so far ahead it makes everything else superfluous. On the other, Europe and China are working precisely to escape that script, building national or open alternatives that prevent anyone — Washington included — from holding the switch.

No one knows the future. But it would not be surprising if, in a few years, the landscape settled into layers like this: an artificial intelligence «good enough for most jobs» that has become free or nearly so, open, downloadable, commoditized — exactly as the cost curve above predicts — and above it a narrow tier of far more expensive models, reserved for frontier tasks: advanced scientific research and, above all, military and national-security uses. These last models would not be sold like a subscription but exported the way weapons are: with licenses, usage restrictions, clauses on where and how, and deliberately weakened versions for foreign partners.

If that were to happen, today's race for the «trillion-dollar model» would look, in hindsight, like a snapshot of an intermediate moment: one in which baseline intelligence was not yet widespread enough to be free, nor strategic enough to be treated as defense material. A moment in which, for a handful of years, it seemed a single model could be worth an entire industry, before the value migrated, as it almost always does, elsewhere: downward into what becomes free infrastructure, and upward into what becomes a state secret.

One figure ties the two worlds together: building the infrastructure now costs as much as the «intelligence» that runs on top of it. A large data center is worth on the order of a billion dollars — a 100 MW facility costs between $0.9 and $1.5 billion — a figure of the same order of magnitude as what it takes to train a frontier model: between $200 and $500 million today, heading toward a range of $1 to $3 billion by 2027.¹¹ The difference is in volume: data centers are built by the thousands, frontier models by the handful. That is why, in aggregate, infrastructure weighs far more than models: McKinsey estimates a cumulative — not annual — investment of $6.7 trillion in data centers through 2030.¹⁰ Value, in AI as in the cloud, settles into infrastructure more than into any single product.

And as in the cloud, sovereignty is already rewriting the market here too. The cloud has consolidated around three American hyperscalers, which together account for roughly two-thirds of the global market, but fear of the CLOUD Act and geopolitical tensions are pushing Europe toward continental providers, from OVHcloud to the Schwarz Group's STACKIT. According to Gartner, European sovereign cloud spending goes from $6.9 billion in 2025 to over $20 billion in 2027 — more than tripling in two years.¹² It is the most direct precedent for what the sovereignty push and export controls can do to models.

The most uncomfortable parallel remains the one with telecoms and fiber in 2000: capex out of scale, overcapacity, price collapse. The survivors, Cisco among them, became solid, profitable companies — but never again close to their 1999 multiples.

It is more prudent to expect, over time, multiples closer to those of mature software infrastructure than to those implied by a winner-takes-all narrative. It is the same pattern as Cisco after 2001: a relevant, profitable company, but never again at the multiples of when it was thought to own the entire network.

Besides, the math today struggles to add up: big tech's AI capex went from about $410 billion in 2025 to about $725 billion projected for 2026.¹¹ And David Cahn's (Sequoia) math, which back in 2024 already estimated around $600 billion in annual revenue needed to justify the spending, is a gap that, with these numbers, is widening rather than closing.¹³

For those leading an organisation, the lesson is practical, not financial. The lasting advantage does not lie in owning access to the most powerful model of the moment, but in proprietary data and in people capable of using any model well. Put more explicitly:

Applied intelligence to evolve human.