The Signal — April 20, 2026
Monday's signal: the silicon wars are getting personal, OpenAI's next model is casting shadows before it arrives, and the ML research world descends on Rio.
Google in Talks with Marvell to Build Two New AI Chips
Google is in discussions with Marvell Technology to co-develop two new chips purpose-built for AI inference, according to The Information. The first is a memory processing unit designed to sit alongside Google's existing TPUs and offload the memory bottlenecks that slow down large model inference. The second is a new TPU architecture built specifically for running AI models in production.
The MPU design could be finalized as early as next year, per Reuters. If the partnership materializes, it puts more pressure on Google's long-running effort to break free from Nvidia's GPU ecosystem. Custom inference silicon isn't new for Google (TPUs have been around since 2016), but pulling in Marvell for dedicated memory architecture suggests the bottlenecks are shifting in ways that general-purpose accelerators can't address.
The bigger pattern: as inference costs become the dominant expense in deploying AI at scale, the companies running the largest models are building their own chips rather than waiting for Nvidia to solve their problems.
Signs Point to OpenAI's Next Major Model in Production Testing
Signals suggest OpenAI's next major model, widely expected to land as GPT-5.5 (internally codenamed "Spud"), has entered production-scale API testing. Multiple observers flagged unusual API behavior on April 19, and reporting from Digit.in indicates that pretraining wrapped up in March 2026. The model is expected to bring native multimodality, handling text, images, audio, and video within a single architecture rather than stitching together separate systems.
Greg Brockman described it as "a significant change in the way we think about model development," per Digit.in. Prediction markets on Polymarket reflect strong consensus around a near-term release, though OpenAI has made no official announcement. None of this is confirmed, and production testing doesn't guarantee an imminent launch — but the convergence of API activity, market sentiment, and insider commentary paints a consistent picture.
If the multimodal integration is as deep as suggested, it would be a real departure from the bolt-on approach that has defined most frontier models so far.
ICLR 2026 Kicks Off This Week in Rio
The International Conference on Learning Representations opens April 23 in Rio de Janeiro, running through April 27. ICLR sits alongside NeurIPS and ICML as one of the three premier machine learning research conferences globally, and this year's program is stacked. Microsoft alone has over 150 accepted papers, according to Microsoft Research. Apple's ML research division and Google DeepMind are also presenting significant bodies of work.
For anyone tracking where the field is actually headed, not where product launches suggest it's headed, ICLR is the week to watch. The papers accepted months ago often preview the capabilities that show up in products a year later. Expect results across reasoning, efficiency, multimodal learning, and alignment dropping throughout the week.
Between custom silicon partnerships, models entering the pipeline, and a week of research presentations, the next seven days should offer a clearer picture of where AI development actually stands as 2026 hits its midpoint.
On the Editor's Desk
Stories we tracked but didn't run today: Alibaba's "Happy Oyster" world model (April 16, too stale for a daily), the ongoing DRAM shortage timeline (covered yesterday), and Cerebras re-filing its IPO (covered in March). Sunday news cycles are thin, but the week ahead looks busy.
