For two years, the AI debate has come down to one binary question: revolution or bubble? Each camp has its charts, its anecdotes and its absolute certainty.
It's the wrong question. Nobody knows the answer - not even the people selling the technology. The right question, for an executive, is far more down to earth: what stays true no matter which scenario plays out?
As it happens, we now have scenarios with actual numbers to look at.
Three scenarios, three worlds
Anthropic's economic institute published a model of AI's impact on the US economy through 2030, in three scenarios. These are numbers from a player with an interest in AI mattering: read them as orders of magnitude, not prophecies.
- Moderate - US GDP reaches $34.1T in 2030, i.e. +1.6% versus a no-AI trajectory. AI is one more good technology, taking a generation to diffuse.
- Substantial - $36.3T, with AI able to do half of cognitive work. One reading of this scenario, summarised by the Darden School (University of Virginia), pairs it with up to 8% annual growth and +4 points of unemployment among knowledge workers. Strong growth and social damage, at the same time.
- Extreme - $44.4T, 15% growth per year, labour's share of national income falling to 45.2% and white-collar wages down more than 10%.
Interesting detail: asked in a survey of nearly 11,000 Americans, people bet on the middle one - median answer around +8.6 to +10% of GDP and ~4.6% unemployment in 2030. Neither apocalypse nor non-event.
And humility is in order: Daniel Kokotajlo, author of the "AI 2027" scenario that popularised the idea of an imminent intelligence explosion, has pushed his median forecast back to 2030. When the most aggressive forecasters revise, nobody should build a strategy on a date.
"Transformative AI", to be precise
The term needs a definition, otherwise it means nothing. We're not talking about an AI that writes your emails. We're talking about a system able to autonomously execute a substantial fraction of an organisation's cognitive work - the threshold is economic, not technical.
Concretely, that system is not a model. It's a model plus a harness: a loop that acts (try → check → retry), tools to touch the real world (your databases, your software), a memory, permissions, and a human who decides. All of it plugged into your company's context. That's exactly what we describe in what surrounds the model: the piece everyone stares at - the model - is the least differentiating one.
The money has already bet on extreme
Google, Amazon, Microsoft and Meta plan $725B of investment in 2026, up 77% year on year, per Tom's Hardware's count. Nvidia is reportedly negotiating more than $750B in deals according to Axios, including backing for OpenAI's compute leases - circular structures that feed bubble fears.
On the other side, the selection is brutal: 61% of global venture capital went to AI in 2025 according to the ValueAddVC tracker, but in the single month of February 2026, 83% of the $189B invested went to three companies. And nearly 40% of AI startups launched in 2024 have reportedly already shut down in under 24 months.
What this means for you: don't tie your organisation to one vendor. Being able to swap models or tools overnight, without rebuilding anything, is no longer an architect's luxury - it's risk management. That's the whole point of portability and open source models.
Meanwhile, in real companies
The contrast is striking. According to the MIT NANDA study, 95% of generative AI pilots produced no measurable impact on the P&L, despite $30-40B in spending. Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027. And McKinsey measures 88% of organisations using AI… for only 6% getting a significant effect on their bottom line.
In other words: the scenario doesn't land the same way on everyone. Some companies are already living in the "substantial" one and others in the "moderate" one - same sector, same models.
You don't pick the weather, you pick the hull
This is where an image helps. You can't decide what the weather will be in 2030. You can decide which hull you put to sea in. A good hull is useless on flat water - and saves you in a storm. The test is simple: what varies here is the intensity of the change, and the same foundations hold at every intensity.
What holds in all three scenarios, with no regrets:
- Your context gathered and queryable - a company brain: decisions, procedures, customer history, what's known on the ground. Useful even if AI plateaus; vital if it accelerates.
- Agents plugged into it, with clean permissions and humans in control of the decisions that matter - the definition of an AI-native company.
- Portability across models and vendors, so you don't depend on any single bet.
- Measurement in euros, not in pilots. A pilot that doesn't move a single line of the P&L belongs to the 95%.
And the social variable deserves to be faced head on: in 2026, 205,000 job cuts were explicitly linked to AI, cited as a factor in 54% of plans according to the layoff counts published this year. The companies that come out best won't be the ones cutting fastest, but the ones that redeploy their people toward judgement, relationships, decisions.
What we do
At Elephant in the Room, we don't sell a scenario. We build the foundations that hold in all three: a company brain that holds your context, business agents that plug into it with guardrails, multi-model orchestration with no vendor lock-in, and you at the helm. And we do it on ourselves first: our own company runs on its brain, our agents work on it through the night. We don't sell a theory we don't apply.
You don't choose the scenario. You choose to be the company that holds in all three.
