AI news ·
Compute access is the primary bottleneck for new AI labs, says Anjney Midha
Compute access is the primary bottleneck for new AI labs challenging industry giants. Developers must use a four-step framework and verifiable feedback to improve models.
AI development is consuming vast amounts of physical and financial resources, reshaping global infrastructure to support the race for artificial general intelligence. Anjney Midha, founder and CEO of compute grid company AMP and the first investor in Anthropic, warns that access to compute infrastructure is now the primary bottleneck for new labs trying to challenge industry giants.
The resource race and compute bottlenecks
Major technology companies currently lead the AI sector largely due to their financial capacity to fund massive infrastructure projects. "If you say you absolutely have to be the first to invent AGI, then you can justify any amount of spending on earth," Midha said. This pursuit has led to the widespread repurposing of critical assets. "Everything from access to electrical grids, GPUs, energy turbines, talent, and even residential real estate are being repurposed to make more and more advanced AI," he said. Midha argues that compute access must not become a permanent barrier to entry, noting that a new lab with strong ideas should be able to secure the infrastructure needed to innovate.
Efficiency and the model development recipe
While the race for AGI dominates headlines, Midha emphasized that AI development spans multiple frontiers. He pointed to Anthropic as a benchmark for sustainable growth, contrasting its methods with the sprawling operations of larger tech firms. "Anthropic is clearly a role model for the rest of the community on how to do it in an efficient way," he said. Developing frontier models follows a structured, four-step framework: pretraining, mid training, post training, and a continuous feedback loop. This predictable workflow allows developers to systematically improve model capabilities rather than relying on brute-force scaling alone.
Verifiable feedback in software engineering
The effectiveness of AI models depends heavily on the quality of the feedback they receive. Midha noted that progress is fastest in domains where feedback avoids hallucinations, such as predicting material properties, rather than subjective tasks like creative writing. Software engineering benefits uniquely from this dynamic, as Midha observed that coding has a very systematized approach to feedback loops compared to other fields. When developers apply verifiable feedback to check outputs against factual requirements, model capabilities improve dramatically, allowing teams to embed these systems into core business functions for measurable cost reductions. Professionals looking to integrate these workflows can explore targeted resources on the AI Learning Path for Software Engineers to understand how structured loops improve development.
Why this matters for IT and development professionals
Compute constraints and the need for efficiency will define the next phase of enterprise AI adoption. IT and development teams must prioritize building systems with verifiable, objective feedback loops rather than chasing unstructured generative experiments. By focusing on AI for IT & Development workflows that integrate directly into existing engineering pipelines, teams can deliver measurable value without requiring the vast capital reserves of industry incumbents.