It’s been a big month in AI and Silicon Valley, with the headlines clustering around several themes: a backlash to tokenmaxxing, the impacts of regulatory pressures, and major IPOs (including SpaceX and Cerebras). The moment reflects a market shifting from idealizing companies like Anthropic to a more healthy debate about where and when value will accrue.
Let’s break down the news and make sense of what it means for the people building today.
In regards to tokenmaxxing, attitudes have shifted sharply over the recent month. The competition to see which employee can spend the most tokens is finding itself at a crossroads as CFOs see high costs but little measurable impact to justify such a large expense. Uber is the canonical example, burning through their 2026 budget in just four months.
Sentiment has also shifted rapidly when it comes to regulatory pressures. It was only in February of this year that Anthropic refused to grant the US Department of Defence unrestricted access to its model—today, the US has barred foreign nationals from using Anthropic’s frontier models. Access is becoming a geopolitical question; in Europe and India, there is renewed support for sovereign models like Mistral and Sarvam.
On the IPO front, SpaceX’s IPO is a signal that a new, very well funded player has joined the field. Instead of there being one leader in AGI, as there seemed to be a year ago, we now have four roughly equally powerful players each with their own different assets to bring to the table: OpenAI with the largest AI consumer footprint through ChatGPT; Anthropic with the best in class models and harnesses as of today; Google with the most money, IP, and talent; and now a distant fourth—xAI, flush with capital and compute access, with Cursor as the harness.
These leaders are trailed by several others who are not far behind: Microsoft, Apple, and Meta. Then there are the Chinese players, which have the advantage of being both open weight and multitudes cheaper. DeepSeek’s latest model scores within a hair of Opus on SWE-bench at roughly a 1/30th of the price, and the cheapest open-source serving runs closer to 1/100th. When a credible alternative costs a fraction as much, and functions well enough, why not make the switch?
This is all driving to a model ecosystem full of burgeoning competition and many well-positioned players.
The model is not the moat
Which leads me to the big idea behind what we’re seeing happening right now. The general consensus has been that there will be a winner-takes-all scenario, with the maker of the best model crowned king. But that’s not what’s happening at all.
Competition continues to be fierce among the labs, making it difficult for any one of them to emerge as the clear leader. This is propelled by several factors. First, the competitive dynamics are driving down the price of intelligence. Next, models improve quickly (a model at the frontier today will look ordinary within a year) but advances can easily be copied through distillation. Any lead that a lab gains appears to be tough to maintain. Finally, regulatory forces will continue to play a bigger part. As AI becomes a national security liability, governments are becoming more involved in which models are released, where infrastructure is built, and who gets access to it; this slows down and complicates how labs reach customers.
This means the model is not the moat. It never was. Intelligence is not a winner-take-all market. Its frontier is also jagged: there is no single frontier but many, and a different lab will lead in each.
Instead, the real value for most companies will come from the application layer, through their proprietary data and workflows—the specific knowledge that sits within companies and industries, the decision traces and context graphs that can’t be replicated.
This is where the product comes in: the harness plus the model, wired into a specific customer’s environment that drives a business outcome. A product can turn a model into something durable: a workflow, a habit, a distribution channel, a customer relationship, a store of usage data. The model remains central, but it becomes the replaceable engine inside something that is much harder to replace.
The labs are steadily moving into selling products, not just selling models and harnesses. In the B2B context, Anthropic has been the most aggressive with Claude Code, Cowork and most recently Claude Tag. OpenAI is turning ChatGPT into a super app spanning coding tools and agents, while moving into devices and robotics. Google is leveraging its broad application footprint across Android and Google Workspace to embed Gemini. SpaceX is likely to leverage Cursor to build out its own products, starting with coding. Looking ahead, Anthropic and OpenAI have begun hiring subject matter experts (lawyers, finance, healthcare workers); we should expect to see more domain-specific products from them soon.
The labs are also expanding their deployment options. Microsoft and Google already have the deployment advantage with their large cloud services teams. OpenAI and Anthropic have formed joint ventures with private equity firms to deploy models within their portfolio companies. OpenAI has launched its Frontier Alliance program to work with global consulting firms to leverage their muscle to accelerate model deployment.
Infrastructure platforms are already seeing the writing on the wall, updating their offerings to be model-agnostic. Databricks’ new Agent Bricks platform supports model choice across ecosystems, from OpenAI to Qwen. Snowflake’s agent harness is designed for the underlying LLM to be completely interchangeable without losing guardrails—the model can be swapped with a single toggle.
We’re moving out of the winner-takes-all phase and into a moment where access to hundreds of high quality models is about to explode, giving founders more tools to build with than ever before, and at cheaper than ever prices.
Get insights directly to your inbox.
How startups win in this era of AI
More competition in the model ecosystem is good for startups building at every level of the stack. There are certainly opportunities for new model companies, and there will also be an increasing number of opportunities for harness providers, but the largest opportunity continues to be building AI-native applications and products. At the same time, the model providers will continue to expand their product footprint to compete with their customers.
For founders today, this is a moment to consider what parts of your product are defensible and enduring.
As we have written before, product companies have the opportunity to build a context graph for their customers and domains, which if done correctly is an enduring advantage. Neither the model companies nor incumbent systems of record capture the decision traces required to build the context graph. The orchestration layer sees the full picture, and that’s where the real value is created. This is one more reason why value accrues in the product layer, not the model.
In addition to having a point of view on what makes your product enduring, startup founders need to really think through customer and market segmentation. A recent conversation with a friend of mine, a chief AI officer of a large telecom company, helped illuminate this for me. There are two questions that founders need to ask themselves.
First, what do your customers care about most when it comes to AI? First-to-market and early adopters are willing to buy a product that’s still rough around the edges, and invest in co-development. Companies who value control want local models that can be run within their VPCs and have high security standards. Those who are cost-conscious are most likely to be cautious, and will run long trials with multiple vendors before making a commitment.
Second, what’s the nature of the problem your product solves? There are several categories your product might fall into. Horizontal productivity tools, like Claude, that are deployed-company wide. There is an opportunity for startups, like with Glean (an agentic context graph for enterprise decision-making) and our portfolio company Viven (which creates AI digital twins of employees), but you are truly competing with the giants in this category. Next, there are products that help make departmental decisions, mostly focused on corporate functions like finance, HR, and GTM operations. These companies are more often looking for reliable “off the shelf” solutions that offer clear ROI. They want customization but rarely need bespoke builds. Eightfold (HR and recruiting), and Maximor (finance operations) are good examples of this. Finally, there are solutions that respond to strategic, often customer-facing, board-level priorities. These are heavily-customized and bespoke solutions, often for teams that want to be the first to market. These customers are more likely to work with companies like Turing or the AI services ventures that the labs have launched.
Building exceptional products is still hard, even without the rapid shifts of the AI era. Many AI apps that look defenseless may not be suffering from some new law of the model era. They may just not be very good products. For founders building products today, the advice remains the same as ever: choose the right market and customer segment, make an informed bet on the technology curve, build something people actually want, and then make it easy for them to buy it from you.
Most of the value of this phase of the AI era will be built at the product layer. The model was never the product. It was never going to be your moat either.