Daily Update - May 19th, 2026
"Demand Is Going Parabolic, Utterly Parabolic" -Jensen
Today’s menu: Big event in utilities driven by AI, some insane AI video, TPU adoption increases, Fintwit mania on 13F, and WTF is a forward deployed engineer. Huh? Learn something everyday!
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
NextEra: AI-driven utility mega-merger
NextEra Energy announced a $67B acquisition of Dominion Energy, the largest utility acquisition in US history. Per Bloomberg, the combined enterprise value runs to roughly $118B. NextEra is already the largest US utility by market value and the largest renewable-energy developer, and Dominion is the dominant utility in Virginia, home to the largest data center market on earth.
The household side is showing up in CPI. Per @KobeissiLetter, US electricity prices rose 6.1% YoY in April, the eighth monthly print above 5% in the last ten. AI data center load is now visible in residential bills, which is the political dimension that shapes multi-year utility planning and FERC review.
Austin: Could this be a bit like the semiconductor foundry story, where the CapEx bill grew so large over time that competitors had to drop out or get acquired because they couldn’t afford the next generation? But in this case, it’s utilities needing to merge to afford the CapEx investments required to meet AI-driven demand.
Granted, the analogy probably ends there. Utilities are a regulated industry and geographically bound. But clearly M&A shakeout is going to happen to keep up with the huge investments needed.
It’s also interesting to watch hyperscalers go vertical into generation to work around industry slowness and risk aversion. A bit like Elon Musk’s Terafab, no?
Decart Real Time AI with 80% MFU on Trainium
Decart announced a $300M funding round at a roughly $4B valuation, led by Radical Ventures, and shipped Decart Optimization Stack (DOS) 2.0 the same day. They basically have three product lines:
DOS = full-stack inference and training platform. Runs on Nvidia GPUs, Google TPUs, and Amazon Trainium, and delivers 1,600+ tokens per second for agentic inference, versus an industry average of about 200, and full-HD video and world model inference at up to 100 FPS.
Lucy = real-time world model for immersive experiences in commerce, gaming, streaming, and advertising
Oasis = world model for Physical AI applications like robotics simulation
Lucy and Oasis run on DOS.
Per Decart’s release, Lucy2 now runs on Trainium3 at greater than 80% Model FLOPS Utilization. Annapurna Labs VP Nafea Bshara called Lucy2’s 80%+ MFU “exactly the kind of efficiency Trainium was designed to unlock.” WSJ reported that Amazon has already deployed Decart’s tech across Twitch, online retail, and the AWS movie and television studios. We embedded their video right here, so you can see the tech (not a sponsor, we think its cool.)
Good thing is, you can use the tech RIGHT NOW ton Delulu. Try it out. Impersonate your boss’ boss in your next team meeting. Post in comments with outcomes.
Austin: Dude. Wow. Even more AI silicon infra opportunities to support these world models!
Vik: Remember that in the Gavin Baker interview we posted on yesterday’s update, Gavin points out that everybody is sleeping on Trainium! Also 80% MFU is insanely good. xAI Colossus was recently reported to be 11%.
Google, Blackstone launch $5B TPU cloud venture
Google and Blackstone agreed to form a new AI cloud company capitalized with $5 billion in equity from Blackstone, which will hold majority ownership, according to reports from The Wall Street Journal and Bloomberg. The venture will deploy Google’s in-house Tensor Processing Units rather than Nvidia GPUs, and aims to bring 500 megawatts of capacity online in 2027. The business is positioned to compete with neocloud providers such as CoreWeave by offering third-party access to TPU compute. The arrangement marks an expansion of Google’s strategy to sell its proprietary AI accelerators beyond its own cloud platform. (Bloomberg)
Vik: Google is finally letting TPUs out of the walled garden, and CoreWeave just got a direct competitor with Blackstone’s balance sheet behind it (BX can write checks CRWV can’t match). If Blackstone is willing to take majority ownership on TPU-only economics, market already thinks TPU unit economics rival Nvidia’s.
Leopold Mania on Fintwit
Shares of Lumentum and Coherent fell after hedge fund manager Leopold Aschenbrenner disclosed he had sold his stakes in both photonics suppliers. His 13F filing also shows that he is short most big AI chip names including NVIDIA, AMD, TSMC, Broadcom, and others. Traders on X said funds were exiting the names in tandem with Aschenbrenner’s move.
Vik: One hedge fund manager exits and the whole photonics complex catches a 9% cold — that tells you everything about how crowded LITE and COHR got into optics euphoria. I dunno, make of it what you will. I remain bullish AI semi.
Quick Hits
Tenstorrent drawing takeover interest from Qualcomm, Intel. Potential $5B transaction. AI inference land-grab is on. (Bloomberg)
Nvidia teams up with Dell to build AI factories at scale using Vera Rubin at scale. Fun quote: “Demand is going parabolic, utterly parabolic” (Nvidia news)
According to @jukan05 on X, Intel is pressuring customers to buy CPUs built on 18A. 🤷🏽♂️
Qualcomm disables 5G NR-NTN on a new chipset, citing the lack of commercial satellites available for testing satellite-to-device broadband. (Light Reading)
Key Data
Notice the funny y-axis just to fit in memory. Yeah, thats how insane it has been. 👀
TIL: WTF is a Forward Deployed Engineer
Funny watch. Plus, I learned what AI labs are hiring for.