Will the Frontier Model Boom Last?

The frontier model companies are putting up software growth numbers that are hard to ignore. Anthropic has been the most dramatic example, with recent reports putting its annualized revenue run rate in the tens of billions, driven heavily by enterprise demand and coding tools like Claude Code. OpenAI is also growing at extraordinary speed, with annualized revenue reportedly crossing $20 billion in 2025. Google is harder to compare directly because Gemini revenue is mixed into a much larger cloud and advertising business, but Google Cloud is clearly benefiting from enterprise AI demand as well.

That growth is coming from real adoption, not just experimentation. The most mature use case so far is software development. Coding has the right mix of high labor cost, measurable output, repeatable workflows, and technical users who are comfortable adopting new tools. AI coding agents are quickly becoming part of how software gets written.

The next wave is broader enterprise automation: customer support, sales content, contract review, internal search, financial analysis, system updates, and workflows that used to require a person to sit between five applications and stitch the work together manually.

On top of that, despite all the AI noise, most large enterprises are still early in their adoption cycle. McKinsey’s 2025 AI survey found that nearly two-thirds of organizations had not yet scaled AI across the enterprise. BCG has made a similar point from a different angle, estimating that only a small minority of companies are “AI future-built,” while a larger group is still trying to move from pilots to real value.

That matches what we see in the market. Startups and early-stage companies are often operating in a different universe from the Global 2000. Many are already building products, workflows, and operating models around AI. Large enterprises are still working through procurement, security, governance, architecture, data readiness, employee access, change management, and budget ownership.

So the bull case for frontier model providers is easy to understand. If the biggest companies in the world are only beginning to adopt AI, and the first phase of adoption is already producing massive revenue growth, then the laggards activating should mean a lot more spend.

But there is another question that does not get asked enough: is a world dominated by frontier models, consumed mostly as a service from a handful of providers, really the most likely long-term outcome?

I do not think it is.

The first reason is that open-weight and lower-cost models are improving too quickly to ignore. The best Chinese models, including DeepSeek and Qwen, have put real pressure on the assumption that only the most expensive U.S. frontier models can handle serious enterprise work. Epoch AI has estimated that Chinese models have trailed the U.S. frontier by about seven months on average since 2023. That is not the same as saying every open model is six months behind every frontier model, but it supports the broader point: the gap is measured in months, not years.

The second reason is price. DeepSeek’s published API pricing for its reasoning model is far below premium Claude pricing. Depending on the exact models compared, the gap can be close to an order of magnitude. DeepSeek’s reasoner has been listed at $0.55 per million input tokens and $2.19 per million output tokens. Claude Opus 4.8 is listed at $5 per million input tokens and $25 per million output tokens. That is roughly 9x on input and 11x on output.

Those comparisons can change quickly. Pricing moves. Discounts matter. Caching matters. Batch processing matters. Some workloads need Sonnet rather than Opus. Some can use Haiku. Some should not use Claude at all. But the basic point stands: there is already a large cost spread across models that are good enough for many enterprise tasks.

The third reason is infrastructure. The frontier model business is not just a software business. It is a compute, power, and data center business. Every new wave of usage consumes tokens. More agents mean more steps. More steps mean more prompts, more context, more tool calls, more retries, and more output. Reasoning models can increase the token load further because they spend more compute getting to an answer.

That is why the infrastructure side of AI has become so strategic. Goldman Sachs recently projected that U.S. data center power demand could rise from 31 GW in 2025 to 66 GW in 2027. Power availability is becoming a constraint. Grid interconnection is becoming a constraint. GPUs are still a constraint. Data center sites are becoming a constraint.

Anthropic’s recent compute deal with Elon Musk’s SpaceXAI/Colossus infrastructure is a good example of what the market is telling us. Demand is there. Capacity is not trivial. And the companies winning enterprise AI workloads are increasingly competing not just on model quality, but on access to enough compute to serve the demand.

That raises a hard question for enterprises. If AI becomes a core operating layer for the business, should every internal workflow default to the most expensive model available?

Of course not.

Nobody would design a cloud architecture that runs every workload on the most expensive instance type. Nobody would store all data in the highest-cost storage tier forever. Nobody would send every customer interaction to the most expensive human expert in the company. AI will not be different.

Some use cases really do need the best available model: complex coding, high-stakes reasoning, advanced research, and certain legal, financial, scientific, and security tasks. In those cases, a small improvement in accuracy may be worth a large increase in cost.

But many use cases do not. A lot of enterprise AI work is summarization, classification, extraction, drafting, routing, search, transformation, and structured workflow execution. Those tasks may need reliability, governance, context, and integration more than they need the smartest model on earth.

This is where the current market is a little distorted. Frontier models are capturing so much growth because they made adoption easy. They are available through the hyperscalers. They come with strong developer ecosystems. They have familiar brands. They are easy to justify inside a large company. Nobody gets fired for buying Anthropic, OpenAI, or Google.

Ease of adoption wins early markets. But ease of adoption is not the same as permanent architectural advantage.

For the current pattern to hold, frontier model providers need to keep achieving breakthroughs that preserve a meaningful capability gap. They also need an expanding set of valuable enterprise use cases that require that extra capability badly enough to justify the price. History suggests that is a hard position to maintain.

The high end keeps moving, but the middle gets commoditized. We have seen this in compute, storage, databases, networking, analytics, and cloud infrastructure. Premium products still exist. Premium margins still exist. But mature buyers learn to segment workloads. They reserve the expensive tier for the problems that need it. Everything else gets optimized.

That is where AI is heading.

The enterprise AI stack will not be one model. It will be a portfolio: frontier models for the work that needs frontier intelligence, smaller proprietary models for lower-cost production use cases, open-weight models where control, cost, latency, or data posture matter, specialized models for narrow tasks, and local or private inference where economics and governance justify it.

The winning architecture will route work intelligently across those options. That requires more than picking a cheaper model. Enterprises need to understand the cost, quality, latency, privacy, and reliability tradeoffs for each use case. They need to measure token consumption. They need policies for when to use premium models and when not to. They need fallback paths, benchmarking, observability, and a way to keep AI costs from becoming a tax on every business process.

This is the future Cogniware is building for.

We are not betting against frontier models. We are betting against waste. Frontier models are extraordinary, and they will remain critical for the hardest and highest-value work. (Cogniware can also help the frontier labs run their models more efficiently, too.)

But we believe enterprises will manage frontier model usage closely. They will push use cases that do not need frontier intelligence to lower-cost AI solutions. They will care about cost per token, token per watt, latency, utilization, and workload placement. They will want AI architectures flexible enough to take advantage of the next open model, the next cheaper model, the next specialized model, and the next infrastructure option.

The companies that win with AI will not be the ones that use the biggest model for everything. They will be the ones that know when not to.