Why the AI Giants Are Quietly Losing Money And What That Means for Real Estate

Everyone is talking about how powerful AI is becoming. Far fewer people are talking about what it costs to operate.

According to a new analysis from SemiAnalysis, a $200 per month ChatGPT Pro subscription could cost OpenAI as much as $14,000 if a user actually pushed it to its limits. Anthropic's top-tier Claude plan isn't far behind. The same kind of heavy usage could ring up around $8,000 in real compute costs. OpenAI starts losing money on subscriptions once usage climbs above 11%. Anthropic hits its ceiling at around 20%.

Read that again. OpenAI doesn't need users to max out the service to lose money. According to the analysis, it happens at just 11% of theoretical capacity.

This isn't a minor accounting quirk. It's a structural tension sitting at the center of how the largest AI companies in the world are operating right now  and it has real implications for anyone thinking about bringing AI into their association or MLS.

The Math Behind It

Here's what's actually happening.

The big AI “frontier” platforms (OpenAI, Anthropic, Google) built their user bases on flat monthly pricing. A simple number you could point to. Easy to justify, easy to budget. It worked brilliantly for growth. Tens of millions of users signed up.

But AI doesn't run on good intentions. It runs on compute. Servers, chips, energy, infrastructure. And that compute costs money every single time a query is processed. The more complex the task, the more tokens consumed, the higher the tab.

The new generation of agentic AI, the kind that can think through multi-step problems, run research, handle long workflows, can require up to 1,000 times more tokens than a simple prompt. That's not a rounding error. That's a completely different cost structure.

When companies rush to enable heavy AI use without thinking about it, things can go sideways fast. One widely reported example illustrates how quickly costs can escalate. According to an Axios report, an AI consultant claimed that an unnamed enterprise client accidentally incurred roughly $500 million in AI expenses in a single month after failing to put usage limits on employee access. While the specifics of the story haven't been independently verified, it highlights a growing concern among large organizations: AI costs can spiral quickly when usage isn't governed.

The result? The AI companies are now quietly navigating a dilemma. They need usage to justify their valuations. Too much usage erodes their margins. They need you to use it… just not too enthusiastically.

What Smart Organizations Are Already Doing

Here's where it gets interesting for your association or MLS.

The organizations that are figuring this out aren't abandoning AI. They're getting smarter about how they deploy it. The strategy gaining traction: route the right task to the right model. Complex queries go to the expensive frontier models. Routine, repetitive work goes somewhere leaner.

That routing alone can cut costs by up to 95%, according to reporting from the Wall Street Journal. Ninety-five percent.

The insight buried in that number is something we've believed from the beginning: most of the questions your members ask don't require the most powerful AI in the world. They require the right AI. The one that's been built specifically for your knowledge base, your policies, your context. One that knows what "MLS access" means in your market, what your education requirements are, and how to answer a compliance question without hallucinating something plausible-sounding but wrong.

That's a fundamentally different product than a general-purpose model being stretched across a flat subscription.

The "Race to the Top" Has a Price Tag

There's a temptation in this industry to reach for the most powerful, most talked-about tool, or the belief that it will be easy to build it yourself. The one with the biggest name and the flashiest demos.

The analysis this week is a reminder that the race to the top is expensive, and the companies running it are struggling to make the math work even at the scale of millions of subscribers.

At the same time, many organizations underestimate what it takes to build a reliable AI solution of their own. Access to a large language model is easier than ever. Building a system that consistently delivers accurate answers, protects sensitive information, stays current as policies change, and earns user trust is something else entirely.

The barrier to entry has never been lower. Anyone can upload documents into ChatGPT or create a custom GPT in an afternoon. The challenge isn't getting an answer. The challenge is getting the right answer, every time, across thousands of member interactions while maintaining security, accuracy, governance, and trust. Most organizations don't have dedicated AI engineers, data scientists, machine learning experts, or AI operations teams on staff, nor should they need to.

For MLSs and associations, the answer isn't to chase the frontier. It's to be intentional. The question isn't, "Are we using the most advanced AI?" It's, "Are we using AI that works reliably for our members, in a way we can sustain?"

Those are very different questions. Right now, a lot of organizations are only asking the first one.

How We Think About This at Voiceflip

We built Ardi for one reason: to make member support better without making your budget worse.

That required making deliberate choices about the technology stack underneath it. We don't route your members' routine questions through the most expensive AI models in the world. We built Ardi’s architecture to use the right models for the right task; efficient where efficiency is possible, powerful where it matters, always grounded in your specific knowledge base.

That's how we're able to offer something that genuinely works for associations and MLSs.

This isn't a workaround. It's sound engineering. The same logic that's now saving enterprise companies millions of dollars is exactly what we've been applying at the MLS and association level from day one: don't use a $14,000 model to answer a $2 question.

What This Means for Your Planning

If you're evaluating AI for your organization, here's what the week's news should prompt you to think about:

Sustainability matters more than capability on paper. A tool that's theoretically powerful but practically unsustainable (because the underlying economics don't work) isn't a long-term solution. The AI industry is already seeing providers restrict usage, change pricing, and renegotiate terms as the cost pressures mount.

Purpose-built beats general-purpose for your use case. Your members aren't asking your AI about quantum gravity. They're asking about MLS rules, CE credits, policy updates, and how to reset their SUPRA lockbox. A system trained on your knowledge, built for your context, and priced for your budget will outperform a frontier model that barely knows your association exists.

The cost conversation is coming. Whether you're considering AI now or planning for next year, the organizations that will succeed are the ones building with a clear eye on total cost of ownership, not just the subscription price, but the real cost of the compute behind what they're deploying.

The big platforms built something remarkable, but the math behind it is catching up with them. The good news for real estate associations and MLSs is that you don't have to be caught in the same squeeze.

You don’t need the biggest AI. You need the right AI. 

Ardi is Voiceflip's AI assistant built specifically for MLSs and real estate associations. It handles the questions your members ask every day — 24/7, across 25+ languages, on the channels they actually use — so your staff can focus on the work only humans can do. Book a demo to see it in action.