September 27, 2026 — Can a 1996 Laser Beat Copper in AI Racks?

Host0:48

PicoJool says VCSELs have been part of data-center optical connectivity since nineteen ninety-six. Now they want a job in AI. The company is PicoJool. On September twenty-fourth, it announced a twenty seven and a half million dollar Series A. Socratic Partners led it, and Hudson River Trading joined in. That follows a twelve million dollar seed. So total funding sits at thirty nine and a half million. The plan is concrete. Bigger teams and facilities in the United States and Taiwan. Then qualifying and manufacturing its optical parts for hyperscalers and data center operators. Here's why you should care. In a big AI cluster, moving data between processors runs into limits on bandwidth and power. The wiring becomes a constraint on the whole system. Picture a kitchen full of world-class chefs, with one narrow door to the dining room. So how does PicoJool want to widen the door? The part is a VCSEL, a vertical-cavity surface-emitting laser, meaning the light shoots straight up out of the chip's surface. The trick isn't one heroic lane. It's lots of parallel lanes, added together. PicoJool's June announcement listed four lanes of a hundred gig, four lanes of two hundred gig, and thirty two lanes of fifty gig. The microVCSEL configurations include fifty gig NRZ and sixty four gig NRZ or PAM four. The number the company leads with is bandwidth. It says its two hundred gig VCSELs exceed thirty seven gigahertz. And in September, it said early sampling has begun. The chips are gallium arsenide. PicoJool says it's working with WIN Semiconductor and other gallium arsenide foundries to get the two hundred gig parts ready for production. It's aiming at two packaging styles: active optical cable and near-packaged optics. The stated goal is optical links as cheap, compact and manufacturable as copper. That's the whole bet. And right now it's a goal, not a measurement. Going wide has a catch. You need a lot of parallel optical lanes, so the manufacturing story has to be real. Founder and CEO Al Yuen says reaching two hundred gig per lane proved the performance is there, and the focus now is execution. Which is refreshingly honest. Execution is the part that eats startups. And here's the detail I love. According to the company, Yuen was releasing VCSEL products back in nineteen ninety-six, starting with gigabit Ethernet. So the freshest pitch in AI networking comes from someone who was releasing VCSEL products when a gigabit was the exciting part. Thirty years of overnight success. A skeptic will want independent qualification data. They'll want a like-for-like comparison with competing links. The announcement has neither, because the products are still being prepared for customer qualification and production. So watch for the real numbers. Power per bit, link reach, error rate, and total system cost against copper. Everything PicoJool is developing is about getting bits from one place to another. The next team has the opposite problem. Their bits are already public, permanently, and they'd like them to stop talking.

4:25

The pitch fits on a sticky note. No soft forks, no operators, no bridges. That's how the group [[alloc] init] sums up Shielded Bitcoin. It's a proposal from researchers Clara Shikhelman, Mikhail Komarov and Aleksei Moskvin. The goal is private transfers on Bitcoin, using the consensus rules that already exist. The problem is familiar. Bitcoin's base-chain history is public. Amounts, timing and links between transactions can be visible, and they can be tied to real people. Here's the mechanism. Shielded value lives in encrypted notes. When you pay someone, you create a new encrypted note for them and publish it inside ordinary Bitcoin transaction data. Two things ride along with it. First, a nullifier for every note you're spending, which is a unique tag. Second, a zero-knowledge proof, a piece of math that proves a statement is true without revealing the details. The authors say the proof shows three things. You're allowed to spend. The inputs exist. And the amounts balance. It never reveals which notes you spent, or how much was in them. Double spending gets stopped by that nullifier. Indexers, the separate software tracking the shielded state, check that each nullifier has never appeared before. Think of a bouncer who knows a ticket was used, but not whose it was. Bitcoin itself doesn't understand any of this. It just stores the bytes and puts them in order, like a public notary. Anyone can replay that history and verify it, so no single indexer is privileged. What can a bad indexer do? The authors say it can delay, omit or serve stale data to a wallet. That disrupts the wallet, but it can't spend your funds. There's a nice touch on keys, too. Spending keys and read-only viewing keys are separate. So you can show someone some wallet information without handing over spending authority. The authors are candid about the limits. A viewing key may expose everything it can see. And a disclosure can't prove nothing was withheld. Which, to be fair, is also true of every expense report ever filed. Now the caveats, and they're big. This is a research design, not a Bitcoin feature. And the way bitcoin actually gets into and out of the shielded system is saved for a companion paper that isn't out yet. It's a very private room, and the door is in the sequel. Bitcoin protocol developers and independent cryptographers will push on the soundness, the metadata leaks and the wallet experience. The authors themselves say data size and timing may remain visible. So here's the open question. Once someone watches when things get posted, and how big they are, how much privacy is actually left? Shielded Bitcoin is trying to move money where nobody's watching. The next company wants tokenized dollars right in front of the banks and the regulators.

7:34

A tokenized-repo trade involving DRW and Marex. Settled on something called the Canton Network, with a stablecoin called U S D C x as the cash. HIFI says it supported that live trade on Tradeweb. It also says it took part in a DTCC tokenized repo pilot in July. Those are the names it's leaning on for credibility. On September twenty-fourth, HIFI announced a thirty seven million dollar Series A led by Left Lane Capital. Left Lane's Matt Miller joins the board. The traction numbers are all company reported. More than seven billion dollars processed a year. Usage among existing customers up more than fourfold in six months. And customers on the platform have onboarded more than ten thousand businesses and two hundred thousand individuals. So what does HIFI actually build? APIs that connect a company's payment workflows to stablecoins and tokenized money. The products cover cross-border payments, money coming in, money going out, and controls. The pitch is that you don't rip out the old rails. You bolt the new one on. Concretely, HIFI says developers can pair its stablecoin settlement with Visa Direct for push-to-card payouts. It also says it's integrated with Circle Payments Network. For engineers, that seam is the interesting bit. HIFI says it's building to connect tokenized settlement with existing payment rails. Somebody has to build that seam, and test it. Co-founder Zach Walsh's line is that money is becoming software. Anyone who's shipped software knows that's not purely a compliment. And he's candid about the gaps. HIFI calls custody and control difficult, unsolved problems. Walsh says the company doesn't think every one of them needs solving before institutions start moving. That's the sentence regulators will read twice. The money goes to more regulatory licenses, a bigger team, and a push into cards and capital markets. Banks, regulators and market operators will scrutinize the compliance side hardest. The open question is how much of that seven billion is recurring revenue for HIFI. And how much custody and compliance responsibility it takes on itself. HIFI's argument is that finance keeps building faster pipes and never quite catches up with the paperwork. Hold that thought, because the paperwork is coming for the next two minutes.

10:17

This is Fast Boot. Two sentences a story, and the tax auditor is holding the stopwatch. California taxes software. Numeral points out that California's newly enacted law extends sales and use tax to prewritten software, including software as a service, starting in twenty twenty-seven. Somewhere, a billing team just learned the cloud has a zip code. Numeral's hundred million. Numeral raised one hundred million dollars for AI sales tax software, and says thirty to forty percent of customers' existing exemption certificates fail validation. Turns out your tax exemption is also exempt from being valid. Bitcoin's rejection inbox. Your Shielded Bitcoin wallet tries to open every new encrypted note with its viewing key, and most attempts fail because they belong to other people. It spends all day getting rejected, which makes it the most relatable software in crypto. Wall Street's Wednesdays. HIFI's founder points to nineteen sixty-eight, when the New York Stock Exchange closed on Wednesdays because trading volume had buried its back offices in paperwork. The original scheduled maintenance window, and nobody even got a status page. If those exemption certificates keep failing at that rate, somebody's tax team is about to rediscover the Wednesday off. Everything so far has been lasers, ledgers and paperwork. The last story is one gene, and it's the one I'd most like to see pan out.

11:53

Under one percent. That's how many untreated human pancreatic duct cells drifted, on their own, toward acting like beta cells, the cells that make insulin. Silence a single gene, and WIRED reports that share rose to about eight and a half percent. The gene is called A L D H three B two. The study, in Science Translational Medicine, reports that loss of its function can cause some duct cells to adopt a beta-like state. WIRED says the work was led by Jian Li, a postdoctoral researcher at Harvard Medical School. And the mouse result is the part that makes you sit up. The team transplanted the altered human cells into diabetic mice. Human insulin circulated. Glucose fell to near-normal levels, and the effect lasted six weeks. How did they find the gene? A genetic screen. They disrupted genes across the whole genome and watched which ones nudged duct cells toward the beta-like state. It's flipping every breaker in the house to find the one that turns on the porch light. Here's why that route matters. This one tries to change the function of duct cells already sitting in the pancreas, rather than first generating replacement insulin-producing cells outside the body, and give them a new job. Now, calibrate. Eight and a half percent is not a wholesale conversion. Roughly one cell in twelve took the job offer. The other eleven kept their day jobs. WIRED's headline says pancreatic cells are one genetic tweak away from treating diabetes. That's forward-looking framing. These were cells in a dish and diabetic mice, not people. There's a real engineering problem, too. WIRED notes that A L D H three B two is used by many cell types. So any therapy would have to hit only the intended cells. Li says the next step is either gene therapy, or small molecules that inhibit the gene. On how the gene actually drives the change, Li's own words are that's the part we need to verify first. Diabetes researchers and clinicians will want durable benefit, and safe, pancreas-only targeting, before anyone calls this a therapy. So that's the open question. What's the mechanism, and can you aim it?

14:18

Today was parallel lasers from PicoJool, a fork-free privacy plan for Bitcoin, stablecoin rails from HIFI, and a gene that nudges pancreas cells toward insulin. Every source is in the show notes, sorted by segment. If this was useful, follow The Daily Reboot wherever you're listening, so tomorrow's episode just shows up. I'm off to try every key on my keyring. See you tomorrow.