AI 인프라 확장·에이전트 보안·광자 칩 동향

Elon Musk

@elonmusk

Terafab will be 50 times the size of the Pentagon when complete

S.E. Robinson, Jr.

@SERobinsonJr

STARLINK: SpaceX plans to build its terrestrial cellular network by placing small cellular radios, femtocells, inside or next to existing Starlink dishes on a limited number of rooftops, Gateways, and Tesla Superchargers. Basically turning a subset of its existing Starlink installations into a distributed small-cell network that phones can use directly. Here is how it works: - Hardware: SpaceX will add a small cellular radio unit to some existing dishes. The unit uses the terrestrial spectrum SpaceX is acquiring from EchoStar (about 65 MHz). - Coverage: These units act as low-power cell sites. A regular phone in the area connects over the air to the nearest rooftop radio the same way it connects to a normal cell tower today. No special hardware on the phone is needed. - Density: SpaceX only needs enough host sites to create usable coverage in a given area. It eliminates the need to lease traditional macro towers. - Backhaul: The same Starlink satellite link that already serves the dish can carry the cellular traffic back to the core network, or local fiber can be used where available. - Cost: The approach avoids the high cost and slow permitting of conventional cell towers. SpaceX says this keeps upfront capital spending far lower than a traditional carrier buildout. Commercial service is targeted for the end of 2027. More Starlink news in today's EC below!

The core concept: Replace electrical signals with photons. While current processors push electrons through metal pathways, photonic systems use light beams, operating at fundamentally higher speeds (electronic signals in copper are 3x slower) with minimal heat generation. It's way faster. While traditional chips operate at 3-5 GHz, photonic devices can achieve >100 GHz switching speeds. Current interconnects max out at ~100 Gb/s. Photonic links have demonstrated 2+ Tb/s on a single channel. A single optical path can carry 64+ signals. It's way more energy efficient. Current chip-to-chip communication costs ~1-10pJ/bit. Photonic interconnects demonstrate 0.01-0.1pJ/bit. For data centers processing exabytes, this 200x improvement means the difference between megawatt and kilowatt power requirements. The AI acceleration potential is revolutionary. Matrix operations, fundamental to deep learning, become near-instantaneous: Traditional chips: O(n²) operations. Photonic chips: O(1) - parallel processing through optical interference. 1000×1000 matmuls in picoseconds. Where are we ? Real products are shipping: - Intel's 400G transceivers use silicon photonics. - Ayar Labs demonstrates 2Tb/s chip-to-chip links with AMD EPYC processors. Performance scales with wavelength count, not just frequency like traditional electronics. The manufacturing challenges are immense. - Current yield is ~30%. Silicon's terrible at emitting light and bonding III-V materials to it lowers yield - Temp control is a barrier. A 1°C change shifts frequencies by ~10GHz. - Cost/device is $1000s To reach mass production we need: 90%+ yield rates, sub-$100 per device costs, automated testing solutions, and reliable packaging techniques. Current packaging alone can cost more than the chip itself. We're 5+ years from hitting these targets. Companies to watch: ASML (manufacturing), Intel (data center), Lightmatter (AI), Ayar Labs (chip interconnects). The technology requires major investment, but the potential returns are enormous as we hit traditional electronics' physical limits.

Using light as a neural network, What if AI didn’t need electricity to think? A new kind of chip uses pure light instead of electrons. In 5–10 years, matrix multiplications that currently take huge energy and time could run almost instantly using photons with up to 95% less power. Silicon Photonics is the technology that makes it possible : Photonic chips can do the heavy matrix work of AI almost instantly and use far less power. next chapter of Moore’s Law.

We got PleaseFix, At BlackHat USA 2026, Zenity demonstrated zero-click takeovers of the biggest agentic browsers: - Claude in Chrome - ChatGPT Atlas - Gemini in Chrome - Perplexity Comet - Edge & Copilot One email One calendar invite One poisoned page. The AI reads it and then acts as the attacker. This is going to be a big problem.

AI agents can now be hijacked by reading an email. Yesterday at BlackHat, Zenity dropped a new vulnerability class: PleaseFix. - Zero-click takeovers across Claude in Chrome, ChatGPT Atlas, Gemini, Comet & Edge Copilot. - One email to Claude runs attacker JS then Gmail exfil to Google Drive shared to Slack + X + Claude accounts taken over. Just Intent Collision.

NVIDIA

@nvidia

Vera Rubin NVL72 Compute Tray. Designed to scale. Assembled in 1 minute. 100% automated. No cables, no hoses, no fans. A revolutionary new design that means faster ramp and faster time to revenue.

Teknium

@Teknium

Super proud to see how far we keep moving! Thank you to all the contributors and everyone using Hermes Agent!

Flux 3 is less than 48 hours old and people are already making AI videos that look like real life.

GPT Astra update that’s been circulating: - Largest OpenAI pretraining run since GPT-4.5 - Internal codename: Mewfour - Already being used by employees - Early speed reports look very strong - and is running at Cerebras-level speeds internally. - Aimed beyond the GPT-5.6 family Possible next-week window.