Why August 2nd will cost some founders more than the next OpenAI release
Your AI model doesn’t matter anymore.
Which country’s servers and which country’s rules you depend on does.
Anthropic shipped Sonnet 5 on July 1st, a Chinese model called MiMo-V2-Pro just became the most-used model on a major marketplace, and neither fact changes your risk exposure.
What changes it: the EU’s next AI Act phase lands August 2nd, 2026, with fines up to €35 million or 7% of global revenue. Founders optimizing for « smartest model » while ignoring jurisdiction are solving last year’s problem. Here’s the harder truth: the model is becoming a commodity.
The map underneath it, servers, chips, rules, is where the real war is being fought.
In 1994, I built my first website in my bedroom, in a housing block near Lyon, on a Packard Bell my mother could barely afford and only bought because I wouldn’t let it go.
Same machine I used to play Dune 2, then Warcraft, for way too many hours.
If you’ve ever played those games, you know the trap every beginner falls into: you obsess over building the strongest unit, the flashiest army. The players who actually win are the ones controlling the territory and the resources underneath, because whoever owns the map decides what you’re even allowed to build next turn.
The entire AI industry is currently obsessing over units while ignoring the map.
The model race is a distraction from the real fight
The model-of-the-month race no longer determines who wins in AI. Anthropic released Sonnet 5 on July 1st, nearly matching its flagship at a lower price, per Build Fast With AI.
Fable 5 returned everywhere after weeks of restriction. Meanwhile a Chinese model, Xiaomi MiMo-V2-Pro, became the most-used model on one of the largest AI marketplaces within a month of most people never having heard of it, according to LLM Stats.
Not because it’s the smartest. Because it offers a huge memory window at a fraction of the price.
That proves the opposite of what most people assume. It doesn’t prove China caught up on raw intelligence. It proves the model itself is becoming cheap and disposable, like electricity. Nobody builds a business strategy around which company sells them electricity.
The real advantage in 2026 isn’t which model you use. It’s which country’s rules and which company’s servers your product depends on for the next five years. Three events this week prove it at three different scales: US-China, France, and Indonesia.
The provocation: if your differentiation strategy is « we picked the smarter model, » you have no differentiation strategy.
Build: what this changes for you
Model-agnostic isn’t a nice-to-have anymore, it’s the baseline. At Agent Nexus, one of our companies, we’ve never debated « which AI model is best », models get swapped in and out constantly, and our product has already run on more than one model without customers noticing.
That’s never where the real risk sits. The real risk is the question most founders never ask until it’s too late: which country’s servers and which country’s rules is this product going to depend on for the next five years? I’ve watched founders in my own network get this wrong the same way, sign with a provider without checking what happens when a new regulation lands, then scramble to rewrite contracts and data-handling clauses under deadline, burning weeks they didn’t have.
Swapping a model costs an afternoon of engineering.
Swapping which country’s rules you live under costs months and legal fees you didn’t budget for.
The provocation: the founders who get burned in the next 18 months won’t be the ones who picked the « wrong » model. They’ll be the ones who never asked which country they were actually building on.
Washington and Beijing are building two separate tech worlds, on purpose
The US and China have stopped pretending they share one AI supply chain. The US tightened export rules in January on advanced AI chip sales to China, according to Oplexa, and China retaliated by restricting exports of the rare materials the US needs, per Informed Clearly.
Both sides are now building parallel supply chains deliberately.
Two separate tech worlds, and the space to move between them is shrinking fast. (we are talking a lot about that on FutureRadar)
Here’s the detail that matters if you’re not in the US or China: smuggling of these chips is common, and Malaysia and Singapore are reportedly used as a back door around the restrictions. Southeast Asia isn’t a spectator in this fight. It sits directly on the fault line, and any company running infrastructure through that region inherits exposure it didn’t sign up for.
The provocation: if you don’t know which side of the chip war your infrastructure provider sits on, you’ve already made a geopolitical bet you didn’t know you were placing.
France just made compliance the cost of doing business
Europe’s AI Act enters its next, tougher phase on August 2nd, 2026 — 26 days out at the time of writing, per ia-info.fr. Fines reach €35 million or 7% of global revenue for banned uses, and €15 million or 3% for non-compliant high-risk systems.
At the same time, the French government announced another €655 million to fund AI development, according to info.gouv.fr. France isn’t trying to out-build the big AI labs. It’s writing tough rules and paying to help its own companies keep up, simultaneously.
A prediction you can check: by August 2nd, more than half of the roughly 200,000 French small businesses already using AI in daily operations still won’t have completed the required paperwork, and the first fines will hit well-known names in healthcare and finance, not the big tech labs. Come back to this after the deadline and see if that held.
The provocation: « we’ll deal with compliance once we’re bigger » is the exact plan that turns a growing company into the example regulators make.
Indonesia is racing to use AI faster than it can govern it
A major AI conference opened in Jakarta this week, and the real question wasn’t which model anyone was using — it was who ends up owning the $10.9 billion in servers and data centers behind Indonesia’s AI push, per CryptoBreaking.
92% of Indonesian workers already use AI tools, the highest adoption rate in the world, according to Introl. Only 24% of business leaders have any real governance in place for that usage. That gap between « everyone’s using it » and « almost nobody has rules for it » is exactly where a new law, an infrastructure deal, or an export ban can blindside a company overnight.
It’s the same map problem I learned as a kid on that Packard Bell: whoever builds the fastest units without securing the territory underneath them usually loses it to whoever was patient enough to claim the map first.
The provocation: 92% adoption with 24% governance isn’t a growth story. It’s a liability waiting for a trigger.
Where the risk actually sits: three regions, one week
| Region | This week’s event | Financial exposure | Deadline |
|---|---|---|---|
| US–China | Chip export controls + rare-material retaliation | Supply chain disruption; Southeast Asia smuggling back-door exposure | Ongoing, no fixed date |
| France / EU | AI Act next phase enters into force | Up to €35M or 7% global revenue (banned uses); €15M or 3% (non-compliant high-risk systems) | August 2, 2026 |
| Indonesia | $10.9B AI infrastructure buildout, Jakarta summit | 92% usage vs. 24% governance gap; ownership of servers still unsettled | No fixed date, moving fast |
The pattern across all three: the exposure isn’t the model you chose. It’s the jurisdiction, the ownership structure, and whether anyone checked the rules before signing.
Why this keeps paying off, or costing you, for years
You can fix a bad model choice in a week. You can’t fix a bad choice of country or provider that fast — it gets paid back over five years, one way or the other. A company that treats « which country, which provider » as a real strategic decision keeps that advantage every time the rules change again. A company that treats it as a technical afterthought pays for that mistake again and again, with interest — the way I’ve watched happen to founders who signed first and checked the rules later.
Same lesson as that Packard Bell all those years ago: the strongest unit loses to the player who owns the map.
If you want to figure out where your business actually stands on this — France, Southeast Asia, or somewhere in between — that’s exactly the kind of conversation we have inside The AI Circle. Or if you want to go deeper on your own situation, grab time with me directly at remybigot.fr/call.
FAQ
Q: Does the AI model I use actually matter for my business risk in 2026? A: Less than most founders think. Models are increasingly commoditized and swappable within a day of engineering work. Your real exposure sits in which country’s servers you run on and which country’s rules govern your data and operations — that’s what takes months, not hours, to change.
Q: What happens on August 2nd, 2026 under the EU AI Act? A: The next, tougher phase of the EU AI Act enters into force. Fines reach €35 million or 7% of global revenue for banned AI uses, and €15 million or 3% for non-compliant high-risk systems, per ia-info.fr.
Q: Is switching AI providers actually risky, or is that overstated? A: Switching the model itself is low-risk and often invisible to customers. The risk is switching which country and legal jurisdiction your infrastructure depends on — that involves contracts, data-handling clauses, and compliance work that can take months under deadline pressure.
Q: Why is Southeast Asia relevant to the US-China chip war? A: Malaysia and Singapore are reportedly used as back-door routes for smuggling restricted AI chips around export controls, per Informed Clearly. Any company running infrastructure through the region inherits exposure to that fault line, even without direct US or China ties.
Q: Are most companies actually prepared for the EU AI Act deadline? A: Evidence suggests no. An estimated majority of the roughly 200,000 French small businesses already using AI in daily operations are unlikely to have completed the required compliance paperwork by August 2nd, 2026, based on current adoption and preparation trends.
Q: Why does Indonesia have such a high AI adoption rate but low governance? A: 92% of Indonesian workers already use AI tools, the highest rate globally, per Introl, while only 24% of business leaders have real usage governance in place. Adoption speed has outpaced regulatory and internal-policy development.
Q: Is being « model-agnostic » actually worth the effort for a small company? A: Yes, and it’s simpler than it sounds. It means designing your product so it doesn’t depend on one specific model, which protects you when providers change pricing, availability, or compliance status. The harder, more valuable work is auditing which country and provider your infrastructure sits under — that’s where real financial exposure hides.
The verdict
Stop asking which AI model to use next quarter. Start asking which country’s servers and which country’s rules your business will still be standing on in five years — because that question, not the model, is the one with an €35 million price tag attached to getting it wrong.
3 questions to ask yourself this week:
- Where does your business actually run — not which model, but which country’s servers and rules? If over 30% of revenue depends on one provider in one country, what happens to your margins if that country’s rules tighten in the next 12 months?
- If your main AI provider got cut off from a market tomorrow — a chip ban, an export restriction, a fine — how fast could you switch? « I don’t know » is already a risk you’re carrying.
- Have you put a number on what a bad choice of country or provider could cost — legal fees, lost weeks, rewritten contracts — before it happens? If that number doesn’t exist in your planning yet, it’s the first one to work out this week.