AI Hardware 2025: Photonics, Analog Compute, and the Green Shift
By Dr. Hernani Costa — June 11, 2025
Emerging photonic interconnects, AMD's analog compute push, and open-source carbon scheduling tools reshape AI infrastructure
Good morning!
Giant models grab the headlines, yet the plumbing beneath them is changing even faster. Today’s issue unpacks three strategic—but still under-the-radar—hardware moves that could tilt the AI race. Let’s jump in.
Lead Story — When Light Meets Silicon: The 2025 Hardware Reset.
Two announcements in the last couple of weeks point to a future where electrons step aside for photons and analog computation:
-
Lightmatter’s photonic interposer and chipset. The $4.4 billion startup revealed silicon-photonics parts that shuttle data between AI dies with light instead of wires, promising bandwidth leaps and major energy cuts. The interposer ships in 2025, and the chipset in 2026.
-
AMD snaps up Untether AI’s team. The stealthy deal adds analog-compute talent and IP to AMD’s growing arsenal, complementing its recent photonics buys and signaling a push to offer low-power inference chips that rival Nvidia’s edge lineup.
Why it matters
- Bandwidth ceilings are near. Training clusters already choke on I/O. Photonic layers deliver orders-of-magnitude headroom without a megawatt bill.
- Latency is user experience. Faster die-to-die hops make agent chains snappier—think voice AI without awkward pauses.
- Power is profit. Analog compute (Untether’s forte) slashes joules per inference; perfect for laptops, kiosks, and anything off the data-center leash.
- Vendor chessboard shifts. AMD’s photonics + analog stack narrows Nvidia’s lead; early adopters may lock in better price-performance before the herd arrives.
Bottom line: If you spec hardware for 2026, photonic lanes and analog tiles should already be on your BOM radar.
Quick Takes
- EU AI Act Energy Clock Ticks—Maybe. Brussels is considering a delay on strict energy-impact disclosures, but officials insist reporting frameworks will still land by summer 2026.
- Speedata Raises $44 Million Series B. The Israeli startup’s “Data-Flow APU” claims it can replace server racks with a single chip for analytics workloads.
- Arm’s First AI Chip Division. Backed by SoftBank, Arm targets a 2025 prototype of its own accelerator, eyeing mass production next fall.
- Nvidia’s Cheaper Blackwell for China. A bandwidth-capped variant dodges export limits, hitting OEMs at $6.5–8 k per GPU.
Tool Highlight — Carbon-Aware Scheduler (Open Source)
Need an emissions scorecard before regulators ask? Drop this MIT-licensed service into Kubernetes, tag low-priority jobs “eco,” and it automatically reschedules to cleaner grid hours, cutting Scope-2 CO₂ by ~30 % in pilot tests. No code changes, just greener compute.
Let’s Wrap Up.
Photon links, analog logic, and carbon ledgers are moving from lab demos to PO numbers. Question: Which hardware angle—optical, analog, or greener scheduling—hits your roadmap first? Reply and let me know; your insights drive our next deep dive.
Until tomorrow—keep the photons fast and the watts low, — First AI Movers Pro
Author: Dr. Hernani Costa — Founder of First AI Movers and Core Ventures. AI Architect, Strategic Advisor, and Fractional CTO helping Top Worldwide Innovation Companies navigate AI Innovations. PhD in Computational Linguistics, 25+ years in technology.
Originally published at First AI Movers under CC BY 4.0.