Compute & Chips

The decision · Map every AI tool and model your company uses to the chip it runs on within 30 days. If more than 70 percent of your AI spending goes to one company, start building a backup plan before that company's next acquisition changes your options.

The decision · If you hold Marvell or are considering a position, watch the August 27 earnings call for two signals: whether management raises interconnect guidance above 70 percent year-over-year growth, and whether it provides a Google revenue ramp timeline. The $12.2B warrant is a seven-year statement about where Marvell expects to sit in the custom silicon market, but the stock at 54x forward earnings is priced on that future, not on today's revenue. Make the call after the guidance, not before it.

The decision · If your organization buys inference compute, track the DX-1 tape-out this year. Olix skipping HBM means its supply chain does not compete with every hyperscaler bidding up the same memory. If DX-1 ships on schedule, it opens a second source for large-model inference at a different cost point than Nvidia. If it slips, the window for alternative architectures narrows. Decide by tape-out whether Olix makes your 2028 procurement shortlist.

The decision · Re-baseline your AI infrastructure plan within 30 days assuming every frontier lab you depend on is building its own silicon. If Anthropic, OpenAI, and Google all ship in-house chips at scale, inference pricing compresses and the value shifts to whoever owns distribution. That means a GPU contract written for 2027 should price in a cheaper open market, not today's shortage. Model your cost per token under both scenarios before you sign the next capacity deal.

The decision · If your organization's inference spend exceeds $200,000 per month, request an Etched benchmark on your own workloads within 90 days. Prefill and decode are different economics. If Etched's low-voltage approach lowers inference cost by even 20 percent on your specific models, the savings compound monthly. Do not wait for public benchmarks. Run your own.

The decision · Re-baseline your AI infrastructure roadmap within 60 days. If your organization is planning GPU procurement for 2027 and beyond, AMD's entry into enterprise AI training changes the pricing equation. The Nvidia premium was built on a single-source assumption. That assumption is now under negotiation. Start pricing dual-vendor architectures. The MI450 does not need to beat Blackwell on every metric. It only needs to be good enough at a price that makes training GPUs a commodity.

The decision · Deployment, not models, is the moat. If you sell an AI product, your risk is that the customer's hyperscaler wraps your workflow into a 'let us help you implement it' bundle. Decide what you own beyond the model call.