The VC Party Is Ending: What Happens When AI Tool Pricing Actually Has to Turn a Profit?

If you have been in tech long enough, you have seen this story before. A wave of new technology arrives, venture capital floods in, prices get subsidized to near zero to grab market share, and somewhere between three and seven years later, most of the players either disappear or merge with whoever has the deepest pockets. Then the survivors raise prices, customers complain, and the ones who built critical dependencies on the cheap tools scramble to figure out what to do next.

We are somewhere in the middle of that cycle right now with AI. And the math is starting to get uncomfortable.

Consider the operating economics. The major foundational model providers are running at significant losses. OpenAI is estimated to be operating at around negative 70% margins. Anthropic is closer to negative 30%. Google is absorbing its AI investment as corporate capital expenditure, which means the losses are real but they do not show up in the same way. These are not small rounding errors. They represent enormous ongoing subsidies that are allowing businesses and developers to access capabilities at prices that do not reflect actual cost.

At some point, the market has to clear. Not every tool survives, and the ones that do will need to charge real prices for real services.

The counterargument is that some players are already turning the corner. Perplexity is running at roughly positive 25% operating margins. Mistral is around positive 15%. These are real businesses with sustainable unit economics, partly because they made choices about model size and specialization that allowed them to run more efficiently. But they are the exception, not the rule. The majority of what is out there is still burning through runway.

At some point, the market has to clear. Not every tool survives, and the ones that do will need to charge real prices for real services. If you are running a business that has built meaningful dependencies on three, five, or eight different AI tools and platforms, you should be thinking seriously about which of those are likely to still exist in two years and what your backup plan looks like if they do not.

The Supply Chain Problem Nobody Is Talking About Enough

There is a less-discussed dimension of this risk that sits upstream of the model providers themselves: the physical infrastructure that makes it all run. Training and running large AI models requires two things in enormous quantities: compute chips and electricity. Both of those inputs are under real pressure right now.

No Helium = No chips and Oil Shortage = Electricity Rationing

The Chips Are Falling:

(Or more accurately: rising dramatically in price — right now as of May 2026)

On the chip side, manufacturing advanced semiconductors requires specialized industrial gases, including helium, at scale. A meaningful portion of the world's helium supply runs through Qatar. When that supply is disrupted, it does not just affect party balloons. It affects production schedules at TSMC and other fabricators that supply the chips powering every major AI system. Missed production quotas have a way of translating into longer lead times, higher prices, and allocation decisions that favor the biggest buyers. If you are a mid-sized enterprise rather than a hyperscaler, you are not at the front of that line.

I deliberately went to Best Buy in Michigan and Media Markt and Saturn in Berlin this month and asked the sales reps one simple question after I bought the newest Google Pixel and Watch and helped my Mom buy an iPad that should last 5-7 years:

The question I asked? ”How fast are phone prices going up?”

MediaMarkt said “About €30/day” Per. Day. Best Buy said “I’ve never seen prices go up this fast; laptops have jumped $400 in just two weeks” and cited the impending oil shortage and rationing due to the Hormuz crisis. This is why I’m urging all friends, family, clients and readers to upgrade to the newest electronics models they can afford because they are about to get much more expensive, and more scarce.

Will The Lights Stay On Or The Bots Be Fed?

(We currently don’t have enough energy production planned to do both)

On the electricity side, communities in Indiana, Virginia, and elsewhere are already experiencing rate increases tied to data center demand. Some of these facilities consume as much power as a small city. The larger ones consume as much power as 900,000 homes.

The political backlash is not hypothetical. Residents, local governments, and regulators are actively debating how to prioritize electricity allocation between industrial AI infrastructure and residential and commercial needs. The answer in several jurisdictions is trending toward protecting residential access first. That creates a real ceiling on how much capacity can be added where it currently exists, and siting new data centers in new locations takes time and capital.

The combination of constrained chip supply and constrained electricity supply means that the cost of compute is not going down linearly the way AI optimists tend to assume. Which brings the pricing problem back into sharper focus. If your input costs are going up and you are already running at deeply negative margins, the math only goes one direction.

Add in objectives like tokenmaxxing: the practice of leadership expecting employees to hit a spending target on AI whether it is useful or not. It is a visible goal that must be met, and some call it “lighting money on fire.” Large model providers that benefit from this practice may end up in better financial shape as a result, at the expense of smaller companies that bet VC money on breakthroughs that did not materialize.

How to Think About This as a Business

The practical question is, “What to do with this information?” A few things are worth considering.

First, audit your actual dependencies. Most organizations using AI tools regularly do not have a complete picture of how many foundational models they are touching, either directly or through the tools they use. That number is often larger than people expect, and the concentration of risk is also often larger. If three of your core workflows depend on a single provider, that is worth knowing now rather than after that provider changes its pricing model or shuts down a product line.

Second, think about price sensitivity in your own stack. The tools you are using at subsidized prices feel like bargains. But if those prices went up by a factor of three to five, which ones would you still pay for? Those are your essential tools. The others are nice-to-haves that might be worth replacing with more durable alternatives before you are forced to.

Unlike employees who ask for raises and vendors and contractors that renegotiate their contracts when they are up for renewal where there can be a negotiation, digital tools can enforce their own raises, on their timeframe. As more work gets offloaded to AI models and agents and companies with their own P&L (profit and loss), the result could become a Faustian bargain from which you can’t escape without shutting down your business when your unit economics go negative.

Third, speaking of unit economics, look for providers with real (and positive) unit economics. Perplexity and Mistral are not the only examples, but they are instructive. Businesses that have built sustainable margin structures are less likely to face sudden pricing reversals or existential crises than those running on fumes and funding rounds. The compliance certifications and trust infrastructure that established players have invested in are also real switching costs that tend to keep those businesses alive longer than pure speculation would.

While it’s tempting to think that much like the steam engine, railroads and electricity, that the cost of AI to the end user will continue to fall — and to some extent that will be true. BUT only after the VCs and institutional investors have recouped their investments. Prices will go up before they come back down when unsubsidized competition is the driving force.

The era of VC-subsidized AI market share is not over yet, but it is winding down. Remember, ChatGPT was launched in November of 2022, so we’re three and a half years out and still no AGI (Artificial Generalized Intelligence). The investment ROI clock is ticking and only has 1.5-3 years left before ongoing investment dries up for lagging tools. Planning for a market where your tools cost what they actually cost is not pessimism. It is just good operations.

Would you like an AI Health Check to uncover areas for improvement in your tech stack and compliance regimen to avoid unwanted surprises? Let's have a conversation!

Book a free exploratory call with Numbers & Letters here: cal.com/eric-boromisa

Disclaimer/Full Disclosure (You made it!): This blog post was generated with the assistance of AI, with N&L human oversight ensuring accuracy and insight. The thoughts and opinions expressed are our own.