what's hot in tech

A note on bias before I start.

I'm a physics nerd and an engineer by education who loves computer science. I loved robotics long before I loved software. This entire framing comes from that lens, and you should take it accordingly.

I'm biased toward deep tech over another AI trends thread. Everyone is already talking about AI, and most of those takes spoil fast. I'm more interested in what has to happen in the physical world that models alone won't fix.

the thesis

AI is eating cognitive work faster than most people's mental models allow. Mine included.

Software, legal work, financial analysis, drug design, code. All of it trending toward absurdly cheap marginal cost.

People keep asking what AI can do next. I'd rather ask where the displaced value goes.

why current AI is a bubble

A lot of prices are baking in AGI. I'm not pricing that in, and I don't think most people pricing it in have a clear picture of why they should.

Scaling returns are flattening. 10x the compute is producing incremental gains. What we have is pattern matching, not reasoning. It's the best autocomplete ever built, but still not cognition. The capex story is shaky: $300B+ in infrastructure, a thin slice coming back as revenue. That reminds me of dot-com capex. Hallucination isn't a bug you patch away. Stochastic generation confabulates by default, and every fix I've seen is duct tape over the architecture.

These are useful tools. I don't see a straight line to AGI from here. From this point on, I want to talk about where I think the money and the hard problems actually are.

where the value migrates

The cognitive layer is getting commoditized. The physical layer is not.

AI can design drugs. Someone still has to manufacture them. It can optimize a factory layout, but someone has to build the robots and wire the place up. It needs compute, and someone has to fab the chips and keep the lights on.

Five themes. In order of how much I think about them.

1. physical AI and robotics

I started here, and I was underestimating it.

The old bottleneck in robotics was cognitive: design, simulation, iteration. Models have compressed that loop. Hardware is still brutal, but the iteration cycle feels maybe 10x faster than it used to. Earlier waves didn't have a 24/7 design engine behind the mechanical teams. The feedback loop really is different now.

I don't know if the useful window is 5 years or 12. I'm pretty sure it isn't 20.

Humanoids are having a moment. Figure AI, Physical Intelligence, 1X, and Tesla Optimus are the names most people know. A few others worth watching: Agility Robotics has Digit handling real tote volumes in logistics. Dexory out of the UK is doing autonomous warehouse inventory and live digital twins. Skild AI is building a general robot brain with ABB, Universal Robots, and NVIDIA targeting factory floors.

Humanoids get the headlines, but physical AI is already deployed. Boston Dynamics, Waymo, Wayve, and Starship are the obvious names in the field now. Eclipse just raised $1.3B for physical AI across robotics, energy, compute, and defense. That says something about where the money is headed.

2. energy

Every other theme in this post depends on solving this one.

Nuclear went from untouchable to inevitable in about 18 months. Tech companies are buying reactors for data centers. SMRs are the default power answer for AI infrastructure now.

Oklo has signed deals with Equinix (500MW) and Switch (12GW through 2044) and just upsized its Aurora reactor to 75MW. Aalo Atomics unveiled a 50MW sodium-cooled reactor purpose-built for data centers. Deep Atomic proposed a combined power-and-cooling SMR at Idaho National Lab.

Is this genuine enlightenment or AI companies throwing money at the problem? Probably both. Reactors are getting built either way.

On fission beyond the SMR story: Blykalla in Sweden is building lead-cooled fast reactors that can burn existing nuclear waste, having raised $50M. Deep Fission is putting reactors a mile underground in boreholes. They have a 12.5GW pipeline and a DOE pilot in Kansas.

On fusion, if the physics cooperates: Zap Energy is using Z-pinch with no magnets and no lasers, having demonstrated 1.6 GPa plasma pressure and raised $330M+. Inertia Fusion raised a $450M Series A in February 2026.

The part everyone forgets is storage and geothermal. Quaise is using millimeter-wave drilling to reach superhot rock. They hit a 100m granite record at 10x prior speed and are targeting a 50MW plant by 2030. Mazama Energy, Khosla-backed, drilled into 331°C rock near Newberry Volcano, the hottest EGS temperature on record. Noon Energy has demonstrated solid oxide fuel cells running 100+ hours with a footprint 20-200x smaller than flow batteries. ION Storage qualified solid-state batteries in the US in March 2026.

Storage is boring. It's also the bottleneck.

3. AI compute

Every layer of the stack feels contested right now.

The Nvidia monopoly story is tired. The challenge is coming from everywhere. Over $1.1B flowed into AI chip startups in a single week in February 2026.

Etched raised $500M at a $5B valuation for transformer-native chips. MatX raised $500M Series B. Ex-Google TPU team, shipping MatX One in 2027, claiming 2,000+ tokens per second on large MoE models. Cerebras raised $1.1B for wafer-scale, the biggest chip you can actually make. Axelera raised $250M for A100-equivalent performance at a sixth of the power, on RISC-V. Tenstorrent raised $693M at $2.6B, Jim Keller's open RISC-V bet against closed stacks. CoreWeave IPO'd at $23B as the GPU cloud that doesn't care whose chip wins.

Inference is projected to be two-thirds of all AI compute by end of 2026. Nvidia optimized for training. These companies are betting the future is inference.

The interconnect story is quieter but real. Copper hit a power wall in 2025. Hyperscalers are putting co-packaged optics and silicon photonics into their AI fabrics. Lightmatter has shown 1.6 Tbps per fiber on CPO chiplets and is pushing an open reference architecture for optical AI infrastructure. The wires are changing anyway, from electrons toward photons.

There's also an architecture wildcard worth watching. Inception Labs' Mercury 2 generates whole sequence chunks in parallel rather than one token at a time: 1,009 tokens per second on Blackwell, roughly 10x faster than GPT-4o-mini on latency, at $0.25/M input, already live on Azure and Bedrock. If diffusion LLMs keep scaling, transformer-native silicon starts to look like a bet on the previous war. Etched just raised half a billion on the opposite thesis. One of them is wrong.

On that note: Llion Jones co-wrote "Attention Is All You Need," then left Google, moved to Tokyo, and co-founded Sakana AI to bet against his own architecture. The company builds nature-inspired AI using evolutionary model merging and swarm-style approaches that sidestep brute-force scaling. If one of the eight people who invented the transformer thinks it isn't the endgame, that's worth paying attention to.

4. biology

The hardest theme to discuss honestly.

De-aging is now an FDA category. That sentence would have sounded insane two years ago.

Life Biosciences got FDA clearance for ER-100 in January 2026 with early human work on epigenetic reprogramming. Retro ran its first in-human lysosomal-function trial in December 2025. Isomorphic (Deepmind's drug design spin-out, effectively AlphaFold 4) has around $3B in deals across Lilly, Novartis, and J&J, though closed models make academics twitchy. Lila raised $550M with Nvidia on the cap table, running autonomous lab loops for antibodies, catalysts, and carbon capture. EvolutionaryScale, from ex-Meta folks, has ESM3 which keeps showing up in serious protein conversations. Colossal is doing de-extinction as a CRISPR engineering program with a story attached.

I hold this theme more loosely than the others. Biology timelines are long and the regulatory path is a mess. Still, the stuff moving through trials now would have sounded like sci-fi a few years ago.

5. brain-computer interfaces

The only technology on this list that makes humans better rather than cheaper.

Neuralink is tripling electrodes to roughly 3,000. Their Blindsight human trials are starting, and that might matter more than the telepathy demos. Synchron has 10 patients, $345M raised, and is the first BCI to natively control iPhone, iPad, and Vision Pro. Blackrock Neurotech and Paradromics both have active patients with distinct architectures.

Synchron wiring into the Apple assistive stack is what got my attention. Once a BCI shows up under Switch Control on an iPad, it stops feeling like a lab curiosity and starts feeling like a product with distribution.

the talk I want to see

The talk I want at Slush isn't another AI keynote. I want an Oklo engineer on what it took to get a 75MW reactor permitted next to a data center. A warehouse operator on what actually changed once Digit started moving totes. A chip founder on why inference silicon looks nothing like training silicon, and why that might make half the current bets wrong.

Or a BCI demo where someone controls an iPad with Switch Control on stage, not a slide about TAM. The model is the easy part of the deck. The permit, the fab line, the tendon, the electrode: that's the talk.