AlphaSense

Interview with an industry expert on how AI is changing the way software gets built ($MSFT, $ANT, $GOOGL): - The expert says Copilot use has grown sharply this year because people now trust AI far more than they did in 2024, when the expert would have written code by hand rather than spent hours checking a model’s output. Models have matured so much that the expert now hands them a proper spec and only validates a few things, pointing to $ANT's Opus and Fable models as well ahead on software intelligence. - According to the expert, employees are now fully using AI to build and keep asking for more credits, because with nobody interested in writing code by hand, running out of budget leaves them idle. Each person gets a monthly budget for Copilot and Claude, while Gemini works on a flat fee with automatic throttling. Unused credits flow to others, and extra use goes through leadership approval. The expert expects this spend to double within a year and possibly reach four times today's level in about three years, if it starts producing revenue. - According to the expert, the return from AI shows up directly in how fast features reach customers. When a client needs a new product capability, the work normally passes through a number of units, which takes many months. With tools like Claude and Copilot, the same team can hand over the architecture and requirements and have features built, tested, documented, and submitted for review in days, cutting delivery time sharply. - The expert emphasizes that $GOOGL has tried to challenge GitHub, building its own coding environment around Gemini, but is nowhere close to displacing it. GitHub has been the source of truth for code for many years, and the volume of code and solved problems committed there is unmatched, which the expert compares to the gap between Chrome and Bing, where people go to the leader for a reason. The expert also highlights that Copilot works as an interface connecting many models, including those from $ANT, $GOOGL, and open-source providers, so developers are not tied to a single one. - The expert sees Claude Code and Claude Cowork as well ahead of any other agentic system for software engineering, mainly on accuracy and on integration at speed and scale, and his main advice to the Copilot team is to copy that approach. He believes GitHub is better placed than anyone to close the gap, since codebases across the globe already sit there, so a Claude-like agentic experience in the Copilot desktop app, even at a higher price per license, would be easier to adopt than bringing in Claude, which still has to connect to GitHub. GitHub could also offer the same capability across models beyond $ANT's.

Interview with an industry expert on why GPU utilization decides whether to build or buy AI inference ( $AMZN, $MSFT, $NVDA ): - The expert says GPU compute is not elastic the way CPU compute is. The cloud providers the expert's team works with always want a three-year contract, and adding GPUs in the middle of a quarter takes at least a month of lead time on top of that commitment. Because of the three-year commitment, the team keeps paying for a GPU instance even after turning it down, so it has had to build engineering optimizations to work around the problem, while CPU capacity can simply be bought on spot instances when demand rises. - The expert sees GPU availability by region as a real challenge, since GPUs are only available in certain regions while customers span hundreds of regions, so routing traffic is difficult, especially when customers have data retention or locality requirements. To manage this, GPU capacity is spread across AWS, Azure, and GCP in selected regions. - According to the expert, $SNOW works with the three largest cloud providers in a rough 60%, 30%, and 10% split, with AWS clearly dominant. $SNOW is described as one of AWS's largest customers, having announced a $6 billion renewal of its spending commitment on its latest earnings call. - On hardware, everything is $NVDA, with the mix chosen on price-performance. Today it is mainly A10G, H200, and B200, plus some L40s. Asynchronous batch workloads open up better optimization opportunities, since they favor throughput over latency, so L40s can replace B200S for much better price-performance. The dollar split across SKUs is dynamic and subject to change. - The expert notes that GPU utilization depends on product usage, which fluctuates a lot. According to him, break-even sits at about 40% average utilization, counting quiet periods like nights and weekends, with business hours running higher. At that level, a company is indifferent between sending inference requests to Bedrock and self-hosting an open-weight model, and anything above it makes self-hosting more favorable, which makes utilization the key lever in the build-versus-buy decision.