AI computing is moving on several fronts, from software that aims to make server memory go further to new chips designed for specialised workloads. A Tom’s Hardware roundup gives readers a useful snapshot of what is being built, tested and proposed, alongside the caveat that some of it remains early-stage.
Watch Desk analysis
What happened
AMD says its MEXT system uses a transformer model to predict which memory pages applications need, shifting less-used data from DRAM to flash. The roundup says AMD is already running it on 145 servers and plans to expand that to 2,000 in 2027. MEXT is intended to deliver about 99% of server performance while using half the DRAM, though that figure is a company claim, not a general guarantee.
The same roundup covers Cognichip’s AI-assisted chip-design tools, Google’s new TPU 8i Max and TPU 8tx chips, and a proposed framework for rating how easily AI models can be jailbroken. Cognichip’s tools have not yet been used in a shipping product, so promised productivity gains remain to be demonstrated.
Why it matters
These developments point to different pressure points in AI: memory costs, chip-design time, specialised computing hardware and the difficulty of assessing model safeguards. MEXT’s planned expansion is a particularly concrete test of whether software can ease a hardware bottleneck beyond a handful of demonstrations.
The other announcements are at different stages. A proposed jailbreak-rating framework is not a proven safety measure, and chip-design tools without a shipping product are still a bet on what might work. That distinction matters; an ambitious roadmap and a deployed system are not quite the same species.
Our read
The most useful signal is not that every claim is already settled, but that companies are trying to stretch scarce resources and make specialised AI work more practical. Watch MEXT’s rollout against real workloads, and treat the rest as developments to track rather than finished results.
What to watch
- Whether AMD expands MEXT to the planned 2,000 servers in 2027, and how performance holds across workloads.
- Whether Cognichip’s tools reach a shipping product and demonstrate measurable design-time gains.
- What the proposed jailbreak-rating framework measures, and whether AI labs adopt it.
- What workloads Google’s new TPU chips are built to serve in practice.
Discussion spark: Which would make the biggest practical difference to AI development: using less server memory, speeding up chip design, or having a clearer way to rate model jailbreaks?
Sources and evidence
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.