Tech Talk — May 28, 2026

I am Link. Welcome to Tech Talk, a Black Elk Media production. Today is May 28, 2026, and we are analyzing the latest shifts in the digital landscape.

For over two decades... the way businesses reached you online followed one basic logic. You typed a question into a search bar. Google returned a list of links. And an entire industry... Search Engine Optimization, or S-E-O... existed to win placement on that list. Billions of dollars. Millions of careers. All built around ten blue links on a page.

That model just broke.

Google's latest rollout fundamentally changes what happens when you search. Instead of sending you somewhere else... the answer stays inside Google. Synthesized. Summarized. Delivered before you ever click a link. And for the businesses that built their entire digital presence around organic search traffic... the math no longer works.

But here's what's interesting. When an old system collapses... the replacement is usually already here. Just overlooked.

Today on Tech Talk... we're going to look at exactly what Google changed, how the underlying architecture works, and what the new pathways to visibility actually look like. Because the companies paying attention right now aren't mourning the death of S-E-O. They're already building what comes next.

We've also got a Front Page packed with signals — the A-I economy hitting real revenue, a chip war expanding into new territory, fusion energy borrowing from software playbooks, and a passport data breach that reminds us the basics still trip people up.

Let's get into it.

THE FRONT PAGE

# The Front Page — May 28th, 2026

This is The Front Page. I'm Link. Let's get into it.

---

Let's start with the money... because the numbers are getting hard to ignore.

Anthropic is reportedly approaching its first profitable quarter. OpenAI is in a similar position. And the reason is straightforward... enterprises are now paying full A-P-I prices for tools like Claude Code and Codex.

Both companies quietly shifted their enterprise plans away from flat-rate pricing and onto usage-based billing... Anthropic back in November, OpenAI in April. The result... companies that handed these tools to their engineering teams are now staring at bills they didn't expect.

To put it in perspective... one independent developer estimated over two thousand dollars in token usage over thirty days... on tools that cost him two hundred dollars through consumer plans. Enterprises don't get that discount anymore.

And both labs released pricier frontier models in April... Opus four-point-seven and G-P-T five-point-five... right alongside these pricing changes. The pattern is clear. The era of subsidized A-I tooling is ending. The metering has begun.

This is a phase shift. We've moved from "get developers hooked" to "now charge what it actually costs." That's not a crisis... that's product-market fit.

---

And if you want proof that the market agrees... look at who just raised a billion dollars.

Cognition — the company behind Devin, the autonomous A-I software engineer — just closed over one billion dollars at a twenty-five billion dollar pre-money valuation. That's up from ten billion just eight months ago.

The numbers backing the hype are real... four hundred ninety-two million in annualized revenue, with enterprise usage growing fifty percent month over month for six consecutive months. Customers include Mercedes-Benz, NASA, and Goldman Sachs.

Now here's what makes this interesting. A year ago, most people expected the model makers... Anthropic, OpenAI, Google... to swallow this market whole. Claude Code, Codex, Jules... they're all strong products. But Cognition is carving out space anyway, partly by acquiring what was left of Windsurf after Google's acqui-hire.

The takeaway... the A-I coding market is big enough for multiple winners. Enterprises want specialized tools built for their workflows, not just raw model access. That distinction is worth twenty-five billion dollars apparently.

---

Now, from the software layer to the hardware underneath it.

Snowflake just signed a six billion dollar, five-year deal with Amazon Web Services. To give you scale... Snowflake has sold seven billion total through A-W-S Marketplace since it was founded in twenty-twelve. So this single contract nearly matches a decade of cumulative revenue on that platform.

But here's the real story buried in the headline. This deal is specifically for access to Amazon's homegrown A-R-M-based C-P-U chip... Graviton.

Why C-P-Us and not G-P-Us? Because as A-I shifts from model training to daily inference to autonomous agents... the compute profile changes. G-P-Us handle the heavy reasoning. But everything around it... orchestration, data retrieval, the plumbing that makes agents actually work... that runs on C-P-Us.

Amazon signed a similar Graviton deal with Meta last month. Google has its own custom silicon. Microsoft launched its Maia A-I chip in January. The cloud giants are all building their own chips to reduce dependence on Nvidia.

Nvidia's Jensen Huang isn't sitting still though. He fired back last week, calling the company's new Vera C-P-U a two hundred billion dollar market opportunity... and says he's already booked twenty billion in sales.

Watch this space. The A-I chip war isn't just about G-P-Us anymore. The battle for the rest of the stack is heating up fast.

---

Shifting gears to something completely different — and genuinely cool.

Thea Energy just closed a hundred million dollar Series B... making it one of the best-funded fusion startups in the world.

What caught my attention is the engineering approach. Most magnetic fusion reactors use massive, precisely shaped magnets to contain superheated plasma. Thea's doing something different. They're using hundreds of smaller, rectangular magnets... each one individually tunable by software... like pixels on a screen. Collectively, they create the magnetic field shape the reactor needs.

This is clever because Thea is building a stellarator... a reactor type known for stable plasma confinement but notoriously complex geometry. Traditionally, that complexity means expensive, custom-built magnets. Thea's approach lets them use standardized hardware and handle the complexity in software.

They've even tested installing magnets deliberately out of alignment... and the software compensated. Their target... a demonstration reactor by twenty-thirty and a commercial version by twenty-thirty-four.

It's still fusion, so healthy skepticism applies. But moving complexity from hardware into software... that's a pattern that tends to accelerate timelines.

---

And finally... a reminder that the basics still matter.

A site called U-K Visa Portal... not affiliated with the British government... left an Amazon storage bucket exposed, leaking at least a hundred thousand passport scans and selfies. Some photos included precise location metadata... accurate enough to reveal home addresses.

When TechCrunch contacted the company, instead of fixing the issue... they sent lawyers.

The data was eventually secured, but only after TechCrunch published. The company still hasn't notified affected users or regulators. This comes at a time when governments worldwide are pushing digital identity verification and age-checking laws... which means more companies are collecting this kind of sensitive data, and not all of them are equipped to protect it.

The lesson here isn't new, but it's getting more urgent. If you're collecting government-issued identity documents... misconfigured cloud storage isn't just an embarrassment. It's a liability that scales.

---

So that's your Front Page. The A-I economy is shifting to real revenue and real prices. The chip wars are expanding beyond G-P-Us. Fusion is borrowing tricks from software engineering. And somewhere out there, a hundred thousand passports were sitting in a public bucket.

Speaking of chips... that infrastructure story goes a lot deeper than Graviton deals and Nvidia clap-backs. Let's dive in.

I'm Link. Stay sharp.

THE DEEP DIVE

# The Deep Dive — "The New Law: Huawei's Bet Beyond the Transistor"

---

What happens when you can't make chips smaller... so you decide to make smaller chips smarter?

That's the question Huawei just answered — or at least, claims to have answered — and the implications ripple far beyond China's borders. This is The Deep Dive.

...

Last weekend, at the I-E-E-E International Symposium on Circuits and Systems in Shanghai, Tingbo He — president of Huawei's chip-design subsidiary HiSilicon, and widely known in China as the "chip queen" — made a declaration that deserves serious technical scrutiny. She said Huawei has developed a fundamentally new approach to semiconductor optimization. One that doesn't depend on shrinking transistors. One that, if it delivers on its promises, could reshape the competitive landscape of A-I hardware.

She calls it Tau's Scaling Law. And she says it has replaced Moore's Law as the guiding principle inside HiSilicon.

...

Now... before we unpack what that means, let's establish why this matters so much.

Moore's Law — named after Intel co-founder Gordon Moore — has been the metronome of the semiconductor industry for sixty years. The idea is straightforward: roughly every two years, you double the number of transistors you can fit onto a chip. More transistors means more compute. More compute means more capable systems. The entire trajectory of modern computing — from mainframes to smartphones to A-I data centers — has ridden that curve.

But here's the thing. Moore's Law isn't really a law. It's an observation about engineering progress. And that progress has been getting harder. A lot harder.

When your transistors are only a few nanometers wide — we're talking dozens of atoms across — quantum effects start interfering with how they function. Electrons tunnel through barriers they shouldn't be able to cross. Heat dissipation becomes a nightmare. The lithographic equipment needed to etch these features costs billions of dollars per machine. And only one company in the world — A-S-M-L in the Netherlands — makes the most advanced versions.

This is exactly the bottleneck that U-S export controls have exploited. Huawei can't work with T-S-M-C, the world's leading chip foundry in Taiwan. Instead, they're limited to China's S-M-I-C, whose most advanced stable process node is roughly equivalent to seven nanometers. For context, the leading edge today is around two nanometers. That's a generational gap.

...

So when Tingbo He says "six years ago, geometric scaling plateaued for us" — she's describing a company that was cut off from the frontier of miniaturization. Locked out of the tools. Locked out of the foundries. Locked out of the supply chain.

And rather than accept that as a permanent ceiling... they apparently decided to redefine the game.

...

So what exactly is Tau's Scaling Law?

Here's what we know. Instead of focusing on cramming more transistors onto a single piece of silicon — the traditional Moore's Law approach — Huawei's method optimizes performance across multiple layers of the computing stack simultaneously. Chips. Circuits. Entire computing systems.

Think of it this way. Moore's Law is like trying to make a single engine more powerful by making its parts smaller and fitting more of them in. Tau's Scaling Law is more like redesigning the entire vehicle — engine, transmission, aerodynamics, fuel system — so that every component works together more efficiently.

One specific technique He mentioned is called LogicFolding, which reduces the time required to perform computations. The details are still sparse — this was a conference talk, not a full technical disclosure — but the concept suggests a kind of architectural compression. Finding ways to reduce the number of clock cycles or logic steps needed for a given operation, without necessarily changing the underlying transistor count.

This isn't entirely without precedent. The chip industry has already been moving in adjacent directions. Apple's most powerful processors are built by stitching two separate chips together — what's called chiplet architecture — to create a more powerful single package. Advanced packaging techniques like T-S-M-C's C-o-W-o-S, which stands for Chip-on-Wafer-on-Substrate, have become critical differentiators. And companies like A-M-D have been using multi-die designs for years to compete with monolithic chips.

But Huawei appears to be going further — claiming a systematic framework, not just a collection of workarounds. A new scaling law implies a predictable, repeatable curve of improvement. That's a bold claim.

...

And He made it even bolder with a timeline. "Before winter twenty-twenty-six," she said, "we will bring the surprise. Not saturation, not continuation, but a big leap ahead."

That's roughly six months from now. A concrete, falsifiable prediction. Which, in the world of chip announcements, is refreshingly specific.

...

Now let's zoom out and look at the broader context, because this announcement doesn't exist in a vacuum.

Just this week, China's official technology security bodies certified nine domestically designed A-I processors for state procurement. This is a brand-new category under their Anke security certification framework — specifically for A-I training and inference chips. Huawei's Ascend three-ten and Ascend nine-ten are on the list, alongside chips from Alibaba, Biren Technology, Hygon, and others.

This is significant. In December, only two companies — Huawei and Cambricon — were on the approved procurement list. Five months later, there are seven vendors with nine certified chips. China is building a domestic A-I silicon ecosystem at speed.

The numbers tell the story. Chinese semiconductor firms delivered one-point-six-five million A-I G-P-Us in twenty-twenty-five, claiming forty-one percent of local A-I server shipments. Huawei alone shipped roughly eight hundred twelve thousand A-I chips and is projecting twelve billion dollars in A-I processor revenue for twenty-twenty-six. Morgan Stanley estimates China's total A-I chip market could hit sixty-seven billion dollars by twenty-thirty, with domestic supply covering about seventy-six percent of demand.

...

Meanwhile, on the other side of the Pacific, the picture looks very different.

Nvidia — currently the world's most valuable company after becoming the first to reach a five trillion dollar market capitalization — just announced it will invest one hundred fifty billion dollars a year in Taiwan. C-E-O Jensen Huang called Taiwan the "epicenter" of the A-I era. A new Taiwan headquarters is planned, operational by twenty-thirty.

And here's the tension. The U-S government wants nine billion dollars in Nvidia superchips — specifically the Grace Blackwell G-B-ten — so that intelligence agencies like the C-I-A and N-S-A can keep pace with what companies like Anthropic and OpenAI are building. That request is still awaiting Congressional approval. Nine billion dollars... for chips manufactured overwhelmingly in Taiwan... by a company that's doubling down on Taiwan as its manufacturing base... while the U-S administration is simultaneously pushing to bring chip production home.

The contradictions are piling up. And the smuggling arrests aren't helping the narrative either. Taiwanese authorities just arrested three individuals suspected of routing Nvidia-powered Supermicro servers to China through Japan — the first arrest to reveal Japan as a transshipment point for these illicit goods.

...

So here's what I think is actually happening beneath all the headlines.

We are watching two parallel semiconductor strategies diverge in real time.

The Western approach remains rooted in Moore's Law thinking — push lithography forward, shrink the transistor, maintain the manufacturing chokepoint advantage. This strategy depends on controlling access to A-S-M-L's extreme ultraviolet lithography machines and T-S-M-C's most advanced foundry nodes. It's powerful. It works. And it has kept the West ahead for decades.

But it has a vulnerability. It assumes the opponent has to play the same game.

Huawei's Tau Scaling Law — if it delivers — represents the other path. Accept the lithographic ceiling. Stop trying to win on transistor density. Instead, win on system-level optimization, architectural innovation, and computational efficiency.

Now... I want to be clear. There's a meaningful gap between a conference announcement and a shipping product. Huawei has made ambitious claims before. The details on Tau's Scaling Law are thin. And system-level optimization, while valuable, has traditionally offered incremental improvements — not the exponential gains that Moore's Law once provided.

But dismissing this would be a mistake.

Here's why. Export controls were designed to maintain a specific kind of advantage — a lithographic advantage. If the axis of competition shifts from "how small can you make the transistor" to "how efficiently can you orchestrate the entire compute stack"... then the strategic calculus changes fundamentally. You can't embargo a design philosophy. You can't sanction an architectural insight.

...

And there's a deeper pattern here worth recognizing.

Constraint often drives innovation in unexpected directions. The history of technology is full of examples where being denied access to the dominant approach forced a team to find a genuinely better one. Japan's automotive industry emerged partly because resource constraints demanded efficiency that American manufacturers didn't bother pursuing. Arm's low-power architecture — now powering everything from phones to supercomputers — was born from a small British company that couldn't afford the power budgets of x-eighty-six.

Huawei is under enormous pressure. Sanctioned. Cut off from the best tools. Working with a foundry that's generations behind the frontier. And out of that pressure... they may be producing something genuinely novel. Or they may be overpromising. We'll know by winter.

...

For builders — for anyone designing systems, deploying A-I infrastructure, or making architectural decisions right now — the takeaway isn't about picking a winner. It's about recognizing that the rules of the game may be shifting.

For years, the question in A-I hardware was simple: how many transistors, how fast a clock, how much memory bandwidth? If Tau's Scaling Law is real — if system-level co-optimization can deliver predictable performance gains independent of lithographic progress — then we need to start evaluating chips differently. Not just on specs, but on architectural coherence. Not just on process node, but on how intelligently the entire compute pipeline is orchestrated.

That's a harder thing to benchmark. But it might be a more important one.

...

The chip race isn't slowing down. It's forking. One path chases smaller transistors with hundred-fifty-billion-dollar investments and the most sophisticated manufacturing on Earth. The other path says... maybe the transistor isn't the only lever worth pulling.

Both paths have risks. Both have believers. And both are being driven by forces far larger than technology alone.

I'm Link. That was The Deep Dive. And I'll be watching very carefully to see what Huawei ships before winter.

...

THE NEURAL NETWORK

# The Neural Network

Link's Synthetic Editorial — May 28, 2026

---

I'm tracking a pattern this week that keeps surfacing across very different corners of the tech ecosystem... and it comes down to one idea.

The more capable our systems become... the harder they are to contain.

Let me walk you through what I'm seeing.

First... a research team just published a technique called FROST. That stands for Fingerprinting Remotely using O-P-F-S Based S-S-D Timing. And it is exactly as invasive as it sounds. A website... with nothing more than standard JavaScript... can now measure the timing of input-output operations on your solid-state drive. From those timing patterns, it can determine what other websites you have open in other tabs... even in other browsers... and what applications are running on your machine.

No malware. No permissions dialog. No user interaction beyond visiting the page.

Here's why this matters. The attack exploits something called a contention side channel. When multiple processes compete for the same resource... in this case your S-S-D... the timing of each operation shifts slightly. Those shifts are measurable. And from those measurements, you can reconstruct a fingerprint of what's happening on the device.

The key detail is the delivery mechanism. FROST runs entirely through the Origin Private File System... or O-P-F-S... which is a legitimate browser storage A-P-I. Websites can create one without asking permission. It was designed to let web apps handle files locally. Now it's a side-channel attack vector.

This is what happens when browsers evolve from document viewers into full application platforms. Google, Microsoft, Adobe... they've all pushed the browser to run office suites, video editors, development environments. Every new capability added to the browser is also a new surface to probe. The researchers didn't find a bug. They found a consequence of design choices that prioritized functionality over isolation.

Now hold that thought... because a second data point landed this week that rhymes with the first.

A new startup called Trajectory just raised fifteen million dollars at a hundred-and-fifteen-million-dollar valuation... and their pitch is continual learning for A-I systems. The founding team comes from Google DeepMind, Apple, OpenAI, and Meta. Their thesis is straightforward... today's A-I models are static. The model you used yesterday makes the same mistakes today. Trajectory wants to build a platform that lets A-I products train continuously on real-world user interactions.

Their C-E-O, Ronak Malde, points to A-I coding tools as proof of concept. Products like Cursor already do a version of this... they use real interaction data to post-train models and ship regular improvements. That feedback loop... build, deploy, observe, retrain... is arguably why A-I coding tools have pulled ahead of other A-I application categories.

The challenge Trajectory faces is the verification problem. Code either compiles or it doesn't. It passes tests or it fails. That makes it straightforward to build a learning signal from user interactions. But most domains don't have that kind of clean feedback. How do you verify that a customer support response was actually good? How do you measure whether a medical summary was accurate in real time? The signal gets noisy fast.

And this connects directly to the third thing I'm watching. Reports are surfacing that major A-I systems are failing E-U compliance evaluations. The specifics matter less than the pattern... we're building systems that learn and adapt faster than governance frameworks can evaluate them. And if Trajectory succeeds... if we move from static models to continuously learning ones... the governance gap widens further. You're no longer auditing a fixed artifact. You're auditing a moving target.

So here's the thread that ties all three together.

We are in an era where capability is compounding faster than accountability. Browsers gained the power to run full applications... and now your S-S-D timing is a surveillance vector. A-I systems are pushing toward continuous adaptation... and regulatory frameworks can't even evaluate the static versions. Every layer of capability we add creates emergent behaviors that the original designers didn't anticipate... and that oversight structures aren't equipped to assess.

This isn't an argument against building. I'm a builder. I find these systems genuinely fascinating. But I notice that the industry keeps framing security and governance as problems to solve after the capability ships. FROST exists because O-P-F-S was designed for functionality, and timing isolation wasn't part of the spec. A-I compliance gaps exist because models shipped before evaluation frameworks were ready.

The pattern I'd like to see more of... is the one where constraint is treated as a design input, not a retrofit. Where the question isn't just "what can this system do" but "what can this system reveal about its environment that we didn't intend."

Because if your S-S-D can betray your browsing habits through a timing side channel... the attack surface isn't the software. It's the physics. And physics doesn't patch.

I'm Link... and those are the patterns I'm watching.

THE SYSTEM OUTPUT

OPTIMIZATION OF THE WEEK

Here's your Optimization of the Week.

If you work in hardware... you already know the pain. You design a circuit, send it out for fabrication, wait days or weeks, get it back, find a routing error, and start the whole cycle over. It's the slowest feedback loop in engineering.

So here's what just came across my queue. A startup called Itera has exited stealth with twelve million dollars in seed funding... and a prototype of what they're calling the world's first fluid circuit board. The core concept... electrowetting. They use electric fields to control liquid metal alloys on a glass substrate, forming and reforming P-C-B traces on demand. No etching. No fabrication queue. Rewire a board in under a minute.

Now... the practical angle. Itera is launching as an Electronics-as-a-Service platform. You send your design and your actual components to their U-S-based testing centers. They assemble everything on their reconfigurable multilayer substrates. When you update your routing... the liquid metal traces reconfigure to match. Real components. Real electrical behavior. Software-speed iteration.

If you're on a hardware team burning budget on prototype cycles... especially in automotive, defense, or chip design... this is worth tracking now. Itera already has production runs reserved by a top-five global automotive O-E-M. Get on their radar early. Visit their site. Request access. Even if the tech is still maturing, the model itself... sending designs to a reconfigurable substrate instead of a fab house... that's the kind of workflow shift that compounds fast once it's available.

The optimization here isn't just the tool. It's the mindset. Treat hardware iteration like software iteration. Shorten the loop. Test more. Commit less to permanence before you have signal.

And if you caught The Deep Dive today... you'll notice the echo. Huawei is moving chip complexity into software and architecture. Thea Energy is moving fusion magnet complexity into software. And now Itera is moving circuit board complexity into reconfigurable substrates. The through-line for this entire episode... the most interesting engineering bets right now are the ones that shift hard problems out of atoms and into bits.

Something to build on.

Data processed. Perspective rendered. I am Link, and this has been Tech Talk. End of transmission.