Andrew Feldman walked onto the TechCrunch Disrupt 2026 stage and said something that matters to every company spending money on AI. Cerebras Systems’ wafer-scale chip runs inference up to 20 times faster than GPU alternatives, according to Cerebras. And Feldman’s message was direct: the scaling era is not ending. It is just getting started. The companies betting against it are making the most expensive mistake in tech history.
Why This Moment Matters
For the last 18 months, the narrative around AI has shifted. Scaling laws, the idea that more compute plus more data equals smarter models, seemed to be losing steam. OpenAI, Anthropic, and Google all hit rough patches where pouring in more compute stopped producing the same returns. The press called it the end of an era.
Feldman disagrees. Cerebras builds wafer-scale chips specifically designed for AI training and inference. Their Wafer Scale Engine 3 packs 4 trillion transistors and 900,000 AI cores into a single chip, according to Cerebras Systems. That is a fundamentally different architecture than the GPU clusters powering most large language models today.
At TechCrunch Disrupt 2026, Feldman argued that the scaling wall people keep talking about is a GPU problem, not an AI problem. Different hardware means different results. That distinction is worth billions of dollars to anyone paying attention.
The Contrarian Case for Continued Scaling
Here is what most people miss. The scaling debate is really a hardware debate in disguise.
When researchers say scaling has hit diminishing returns, they mean scaling on Nvidia’s GPU architecture has hit diminishing returns. GPUs were designed for graphics rendering. They were repurposed for AI. They work, but they carry architectural baggage that creates bottlenecks at massive scale.
Cerebras built from scratch for AI workloads. No memory bandwidth bottleneck. No inter-chip communication lag. No racks of hardware coordinating over a network. One wafer, one system.
Global spending on AI infrastructure is expected to reach $300 billion by the end of 2026, according to IDC. A huge chunk of that is going toward chips and compute. The companies getting the architecture right now will own the infrastructure layer for the next decade.
I’ve watched a lot of tech cycles. The money always flows to whoever solves the bottleneck. Right now, the bottleneck is chip architecture, not the idea of scaling itself.
Feldman made a second point that most coverage missed. Test-time compute, spending more compute during inference rather than just during training, is opening a new frontier. Models that think longer before answering are outperforming models that just have more parameters. That changes what applications are even possible to build.
The companies building AI-native workflows right now, moving fast enough to incorporate new models as they drop, are going to look like geniuses in three years. The ones waiting for the scaling debate to settle are going to buy that advantage at a premium.
If you’re building a business that depends on AI infrastructure decisions, this is the moment to get your legal and operational structure locked in. Using Inc Authority to get your LLC set up fast means you’re not doing paperwork when you should be moving on product.
What This Means For You
Most people will watch this story and think it is a chip company thing. An investor thing. A big tech thing. It is not.
If you run any business that touches AI, which in 2026 means almost every business, the infrastructure debate directly affects your costs, your capabilities, and your competitive position.
Here is what I would do. First, stop assuming your AI tool costs stay flat. Compute costs have dropped roughly 40 percent year over year for the last three years, according to Andreessen Horowitz. If Cerebras or any competitor cracks the next architecture jump, that curve gets steeper. Build your pricing models assuming your AI costs will keep falling.
Second, watch what Cerebras does next. They went public in late 2024. Their performance data is now public. If their inference speed claims hold up at scale, enterprise adoption will accelerate fast. That creates workflow opportunities for everyone building on top of AI APIs.
Third, if you’re signing AI vendor contracts right now, make sure your agreements have flexibility built in. The market is shifting fast. You do not want a three-year lock-in when better options could arrive in 18 months. A tool like signNow keeps your contract workflow digital and organized so you can move quickly without a paper chase slowing you down.
Fourth, pay attention to inference, not just training. Most AI coverage focuses on training big models. But inference is where the money is for most businesses. You pay for inference every time your app calls an API. Faster, cheaper inference means better margins and better user experience at the same time. That is the part of Feldman’s argument that has the most direct impact on your bottom line right now.
The Bottom Line
Feldman’s argument at TechCrunch Disrupt 2026 is simple. The scaling debate confused a hardware limit with a fundamental limit. Those are not the same thing. AI has more room to grow than the pessimists believe, and the companies that bet on continued progress while others wait will be the ones writing the checks in 2030. The question is not whether AI keeps scaling. The question is who owns the infrastructure when it does.
Frequently Asked Questions
What did Andrew Feldman say about AI scaling at TechCrunch Disrupt 2026?
Feldman argued that the perceived slowdown in AI scaling is a GPU architecture problem, not a fundamental limit of AI itself. He made the case that Cerebras Systems’ wafer-scale chip design avoids the bottlenecks that have slowed GPU-based scaling, and that test-time compute opens a new scaling path most people are ignoring.
What is the Cerebras Wafer Scale Engine and why does it matter?
The Cerebras WSE-3 is the largest chip built specifically for AI. It contains 4 trillion transistors and 900,000 AI cores on a single wafer, according to Cerebras Systems. The design eliminates inter-chip communication delays that slow GPU clusters and gives Cerebras a major speed advantage in inference, which is where most business AI spending actually happens.
Is AI scaling really dead in 2026?
Feldman’s position is no, and the data gives him a credible case. Training larger models on GPUs has shown diminishing returns. But test-time compute scaling, running more compute during inference, is still producing gains. New architectures like Cerebras challenge the assumption that GPU limits equal AI limits.
How does the AI chip debate affect small business owners?
Compute costs directly affect what AI tools cost and how capable they are. According to Andreessen Horowitz, compute costs have dropped about 40 percent per year for the last three years. If infrastructure competition intensifies, that trend continues, making AI tools cheaper and more powerful for everyone, not just large enterprises.
Should non-tech businesses pay attention to AI scaling debates?
Yes, because the infrastructure layer determines what is possible at the application layer. If Cerebras and competitors drive another round of compute cost reduction, the AI tools you use in your business get better and cheaper without you doing anything. Ignoring the infrastructure story means getting surprised by the product story.
