He is risen, unlike stocks. Happy Easter!

Google published a research paper last week. The memory chip industry read it, panicked, and shed $95 billion in market value within 24 hours. Bit of an overreaction, you might think, but then you read it.

TurboQuant is a new method for compressing the working memory AI models use mid-conversation by at least 6x, with no loss in accuracy and no retraining required. On Nvidia H100 GPUs, it delivered an 8x performance boost. The paper was produced by Google Research and Cloudflare's CEO has hailed it as "Google's DeepSeek moment."

Samsung, SK hynix, Micron, SanDisk and KIOXIA collectively lost $95 billion following the release, but the stock panic may be the least interesting consequence. If running AI agents suddenly costs a sixth of what it did last week, the economics of digital labour shift overnight, and the comparative cost of human labour becomes increasingly less attractive…

🧐 What's in it for me? Cheaper inference means cheaper AI products. If you use anything powered by a large language model, expect prices to fall and capabilities to grow. The compression war has officially started.

💵 Out of the Lab: TurboQuant is software, not silicon, which means the winners may not be chipmakers. As inference gets radically cheaper, the value may shift to whoever orchestrates AI workflows.

In the 1990s, Russia strapped a 20-metre mirror to a spacecraft and tried to bounce sunlight onto Siberia. The first test produced a moonlight-bright spot 8km wide. The second mirror failed to deploy. The programme was cancelled and for three decades, nobody tried again.

Then, SpaceX brought launch costs from roughly $54,000 per kilogram to under $3,000, and now companies like Reflect Orbital are waging war on the night sky. 

The California company wants to launch thousands of satellites carrying giant mirrors into low Earth orbit, each redirecting sunlight onto Earth after dark over a 5km stretch. The prototype, EARENDIL-1,  features an 18-metre mirror and could launch as soon as mid-2026 pending FCC approval.

The pitch: light up the night and charge $5,000 an hour. Backers include Sequoia and Lux Capital. Critics include all those creatures that call night their day. 

🧐 What's in it for me? Unless you own a solar farm or are very rich and afraid of the dark, this probably won't touch your life directly. But if the concept works, it raises genuinely novel questions about who owns the night sky, and whether that's a market anyone should be selling.

💵 Out of the Lab: Space-based solar is attracting serious venture money. The question is which approach survives contact with orbit.

As a rule of thumb; bigger batteries charge slower. Your phone takes an hour. Your car takes all night, and a grid-scale installation takes the better part of a week. Two centuries of electrochemistry, and nobody has cracked this, but an Australian team just broke the rule using quantum physics.

The researchers built the world's first proof-of-concept quantum battery, charged wirelessly by laser. Instead of storing energy through chemical reactions like a normal battery, it exploits a quantum trick called super absorption, where all its storage units charge simultaneously in a single flash of light. The strange part; add more units and the whole thing charges faster, not slower. Double the size, halve the wait.

The catch, because there is always a catch with quantum anything: it currently stores energy for nanoseconds and holds a tiny charge. Then again, the Wright brothers' first flight lasted 12 seconds. 

🧐 What's in it for me? Their first practical home could be quantum computers, which need exactly this kind of fast, precise energy delivery. If that works, everything quantum computing promises gets closer.

💵 Out of the Lab: Quantum batteries are barely past proof-of-concept, but the first-mover dynamics are real. CSIRO is openly seeking commercial partners, which is unusual for a national lab.

Chronic pain affects 50 million Americans, costs the economy up to $635 billion a year, and the best drug we have for it also rewires your brain's reward system, builds tolerance, and killed 80% of the 600,000 people who died from drug use in 2019. Not ideal.

Now though, researchers have built something different; a gene therapy that reproduces the same pain killing effect without touching the brain's addiction circuitry. Using an AI system trained to read pain behaviour in mice in real time, they designed a molecular switch that targets only the pain circuits morphine acts on. Switch it on: pain drops.

The researchers are calling it the first CNS-targeted gene therapy for pain. It's preclinical, so trials are next, but the blueprint is impressive enough to be published in Nature and the team has filed provisional patents.

🧐 What's in it for me? If you know anyone trapped between chronic pain and the pharmacy roulette of opioid prescriptions, this is one to watch. Still years from a prescription, but the approach is fundamentally different from anything currently available. The drug that replaces morphine might not look like a drug at all.

💵 Out of the Lab: Non-opioid pain is becoming biotech's most crowded gold rush, and the $635bn annual burden explains a lot.

Erythritol, the artificial sweetener packed into protein bars and "guilt-free" everything, may not be that innocent after all. Researchers at the University of Colorado Boulder found that exposing brain blood vessel cells to a single-serving for three hours reduced their ability to relax, increased constriction, and impaired clot-breaking capacity. 

Tighter vessels, stickier blood, higher stroke risk.

This follows a study of 4,000 people linking elevated blood erythritol to significantly higher rates of heart attack and stroke within three years. It's cell research, not a clinical trial, but the direction of travel is clear enough to warrant taking pause. 

Until next time.

Like what you're reading? Share toast with a friend.