$217 Billion for Two Providers: What AI Money Concentration Means for Your Company's Tool Choice

$510 billion. That's how much venture capital flowed into start-ups worldwide in the first half of 2026 - more than in all of 2025 combined, which came to $440 billion according to Crunchbase data. An impressive number in its own right. But the real headline is in the distribution: $217 billion of that, or 43 percent of total global AI funding, went to exactly two companies - OpenAI and Anthropic.

For you as a Mittelstand entrepreneur, that initially sounds like a footnote from the world of finance. It isn't. This concentration has very concrete consequences for which AI tools you'll use to set up your company over the coming years - and how dependent you make yourself on individual providers in the process. An event from June 2026 shows that this risk is not an academic exercise but has already hit real businesses.

It's worth pausing for a moment to let the scale sink in. $217 billion is more than the gross domestic product of many mid-sized economies. This sum flowed within a few months to two privately held companies, both founded only a few years ago. No other economic sector in recent history has seen a comparable concentration of capital in so few players - not the automotive industry, not the pharmaceutical sector, not even the big platform companies in their respective founding phases. For you as a decision-maker, this means: the two companies currently setting the reference models for generative and agentic AI are also the ones with the greatest financial firepower, the greatest influence on pricing and product development - and, as June 2026 would show, still not immune to abrupt, externally forced cutbacks.

Two companies, $217 billion - how it happened

Let's look at the numbers in detail. OpenAI closed a $122 billion funding round in March 2026, reaching a valuation of $852 billion. Among the investors: Amazon with $50 billion, Nvidia and SoftBank with $30 billion each. A few weeks later, in May 2026, Anthropic raised $65 billion in a Series H round and was valued at $965 billion - making it more valuable than OpenAI for the first time. Anthropic's annualized revenue run rate had already exceeded $47 billion by that point.

These sums are historically unprecedented. But they also mean: an enormous share of the capital that could otherwise be flowing into the breadth of an emerging market - into smaller providers, open models, specialized niche solutions - is instead concentrating on two platforms. For market dynamics, that means: anyone setting up an AI strategy today by blindly tying themselves to "the market leader" is, in effect, tying themselves to one of two companies whose financial gravity keeps growing. That might look like the safest choice in the short term. In the long term, though, it also means more bargaining power on the vendor side, more influence over prices, roadmaps, and product decisions - decisions you as a customer don't control.

On top of that comes a psychological effect that's often underestimated in investment decisions: the bigger and better known a provider becomes, the more choosing that provider is seen as a "safe" decision - very much in line with the old IT maxim that nobody gets fired for buying the market leader. But it's exactly this logic that leads many companies to stop asking the real question at all: not "which provider is currently the best?", but "how do I build my processes so I can swap out the provider at any time without it hurting?". The answer to that second question determines your company's resilience in the long run - regardless of how the balance of power between providers develops in the years ahead.

The real test: when two flagship models vanished overnight

What sounds like an abstract concentration risk on paper became very concrete in June 2026. On June 12 and 13, the US Department of Commerce, as part of export control measures, forced Anthropic to take its two new models Claude Fable 5 and Mythos 5 offline worldwide for all users - the reason given was a reported jailbreak security risk. It wasn't until June 30, 2026, a good two weeks later, that this restriction was lifted again.

Imagine your company had firmly and without any fallback tied a central business process - customer service, contract review, content production, code generation - to exactly one of these two models. For two weeks, that process would either have come to a complete standstill or you'd have had to improvise an emergency solution in a hurry. That's no longer a hypothetical scenario - it actually happened in June 2026, triggered not by a mistake on the provider's part, but by a government order that neither Anthropic nor its customers had any influence over.

What's notable here: this wasn't some small niche product, but brand-new models from one of the two by far most capital-rich and highly valued AI companies in the world. If even a provider of this size can be forced to shut down models worldwide from one day to the next, that shows: capital strength doesn't protect against regulatory or geopolitical intervention. It's no substitute for a fallback strategy of your own.

For your company, the crucial lesson here isn't that Anthropic or OpenAI are unreliable partners - both companies communicated quickly and transparently in this case, and the models were ultimately back after 18 days. The lesson is rather structural: even the best possible provider can be interrupted by decisions outside its own control - whether through export controls, geopolitical tensions, regulatory interventions by individual states, or simply an internal company safety decision. Anyone who sets up their operating processes so that a single model becomes a single point of failure passes this external risk straight through into their own business operations, unfiltered. Anyone who instead works with an interchangeable model layer from the start can, in an emergency, switch to an alternative model within minutes instead of weeks - without customers or employees even noticing.

The gap between concern and the ability to act

A Zapier survey of 542 US C-level executives with active, paid AI contract partnerships, conducted between late January and early February 2026, shows that this very risk has already registered with many decision-makers. The result is revealing - and sobering at the same time.

81 percent of surveyed decision-makers are at least somewhat concerned about dependence on a single AI provider. Yet only 6 percent say they could switch their primary AI provider without operational disruption. For 47 percent, a switch would mean at least one core function going down.

These numbers describe exactly the situation many companies - including in the German Mittelstand - currently find themselves in: risk awareness is there. Structural preparation is missing. Anyone who invests today in a single model, a single API, a single vendor interface without inserting an abstraction layer in between is building exactly the lock-in situation that, according to the survey, the vast majority of their own peers explicitly worry about. The gap between the 81 percent who are concerned and the 6 percent who can actually switch is, at its core, an architecture question - not a question of good intentions or market awareness.

The second number from the survey is also interesting: for 47 percent of the companies surveyed, a forced provider switch would mean at least one core function failing. In plain terms: nearly half of companies with active AI contracts have by now built business processes that simply wouldn't work without their current provider. That's not an abstract number from a distant future, but a snapshot of today's status quo - collected in early 2026, months before the Fable 5/Mythos 5 shutdown showed how quickly a theoretical risk can turn into an acute problem. If you're currently building or expanding your own AI-supported processes, now is exactly the right moment to ask whether you're moving toward that 47 percent or toward the 6 percent.

Agentic AI makes the problem worse - when governance is missing

The urgency of this question is growing because how AI is used is fundamentally changing: away from individual prompts, toward autonomous agents that carry out multi-step processes on their own. According to Deloitte's study "State of AI in the Enterprise 2026," 74 percent of companies plan to use agentic AI at least moderately within the next two years. At the same time, however, only 21 percent already have a mature governance model for operating such autonomous agents.

This combination is tricky. A single chat access point that fails is annoying but usually workable. A network of agents deeply embedded via APIs in your business processes - customer communication, quote generation, scheduling, data processing - and hard-wired to a single model from a single company, is a completely different class of operational risk in the event of an outage. It's precisely in this transition phase, in which many Mittelstand companies are building their first agent infrastructure, that the fundamental decision is made that can later only be corrected at considerable cost: do you build on an architecture that can swap out models? Or do you hard-wire your processes to a single provider's API?

Deloitte's numbers can also be read as a time window. The overwhelming majority of companies haven't yet finished building their agentic AI infrastructure - they're right in the middle of building it. That means: the decisions being made right now in governance committees, IT departments, and executive teams will shape corporate IT for the next five to ten years. Governance for autonomous agents means more than just access rights and approval processes - it also has to answer the question of what happens if a central model is no longer available. A governance model that doesn't answer this question is, at its core, incomplete, no matter how mature it appears on other points.

Anyone closing the governance gap Deloitte describes should factor in model independence from the start - not as an afterthought safety net, but as a founding principle of the architecture.

The competition is catching up - and makes the ability to switch even more valuable

Another aspect makes fixating on a single provider economically risky: the ranking of the most capable models is far from set in stone. One day after the Fable 5/Mythos 5 shutdown, on June 13, 2026, the Chinese company Z.ai (formerly Zhipu) released its open model GLM-5.2 with 753 billion parameters under an MIT license. On the FrontierSWE agentic benchmark, GLM-5.2 scored 74.4 percent - only about one percentage point behind Claude Opus 4.8, one of the leading proprietary models. Several sources put GLM-5.2's cost at around a fifth of what comparable performance costs at established providers.

What does that mean for you? If you tie your company irrevocably to a single model today, you're not just taking on a dependence risk, but also a cost risk: you may end up paying indefinitely for performance that's available almost as well elsewhere for a fraction of the price - just because switching would be too technically cumbersome. An architecture that treats models as interchangeable components, by contrast, lets you regularly reassess which provider currently offers the best balance of performance, cost, and availability - and adjust your decision accordingly, instead of making it once and living with the consequences for years.

The timing is also worth noting: GLM-5.2 appeared one day after one of the two dominant US companies was forced to take two flagship models offline. For companies that would have needed an alternative at that moment, a serious option suddenly became available - provided their technical architecture was even able to plug in an additional model. This coincidence of timing illustrates the core point of this article very vividly: the value of a model-agnostic architecture doesn't show up in theory, but exactly in the moment a single provider fails and a quick alternative is needed.

What this concretely means for your tool choice

The conclusion from all this is not to stop using leading models like those from OpenAI or Anthropic - both continue to deliver outstanding performance, and the enormous investment sums are flowing not least into exactly this ongoing development. The conclusion is rather about how you integrate these models into your company.

Three practical points that follow from current developments:

First: avoid hard-wiring a single vendor's API at the process level. An abstraction layer that wraps model calls turns a later switch into a configuration change rather than a rebuild.

Second: regularly test alternative models on the same tasks instead of relying on a decision made once. Precisely because the field can shift significantly within a few months - as the GLM-5.2 example shows - a recurring reality check is worthwhile.

Third: for critical business processes, define an emergency plan in advance for the case of a provider outage - whether due to a technical fault, a price change, or, as happened in June 2026, a regulatory intervention. The question isn't whether such an event will happen again, but when, and with which provider next.

The capital concentration of $217 billion in two companies is a signal, not a reason to panic. It shows where market power in the industry is shifting - and makes clear why the architecture of your AI tools matters more today than the question of which single model currently leads the field. Companies that account for this early won't be among the 47 percent who feel an outage next time - they'll be among the 6 percent who simply keep working.