Tag: AI Security

Weekly Technology Brief #4

The Bottleneck Is Moving Again!

29 October 2026

Dear readers,

A useful way of misunderstanding technology is to stare at the thing everyone is talking about:

With AI, that means the model, while with quantum computing, it is the qubit. What about Space? You guess it: It is the rocket. And in the same fashion, with energy, it has to be the new reactor😊.

This week offered several reminders that the really interesting action is increasingly happening somewhere else: In verification, memory, manufacturing, existing infrastructure and the awkward business of turning impressive technology into reliable systems. Yeah, the bit that nobody is really talking about because it is not, what’s the word for it, sexy!

Of course, all that sounds less exciting than “here is another breakthrough” but face it, the boring is also, probably, the one that is more important.

So, here we go…

1. AI May Be Making Proof Cheaper: Just as It Makes Attack Cheaper!

On 22 October, Anthropic reported an unusually interesting experiment in formal software verification.

Working with Theorem, researchers used Claude to prove security properties at the compiled assembly-code level across 123 real-world open-source libraries. The system produced about 180 million lines of machine-checkable Rocq proofs covering 239 fuzz-test entry points. Along the way it also found bugs that extensive conventional testing had missed.

Five days later came the uncomfortable mirror image as Anthropic reported that an internal research model had discovered and constructed multi-stage exploits against real software including Chrome, Android, Windows and Linux. One Linux vulnerability enabled zero-click remote code execution, while another attack bypassed layers of Android lock-screen protection. The vulnerabilities were disclosed to vendors and patched before details were published. Well, that sound responsible, right?

Now, put the two stories together and something more interesting appears:

AI is becoming better at both sides of the software-security equation:

Find vulnerabilities → Exploit them

but also:

Specify properties → Mathematically prove them

That could change the economics of formal verification. Historically, proof has been enormously expensive: Anthropic notes that the seL4 kernel took roughly two person-years to implement and about twenty person-years to formally verify (I briefly checked these claims and, on the surface, they look legit!). LLMs are particularly suited to producing huge quantities of tedious but proof code. Whether all these codes are mechanically checkable (probably are) and can be fully trusted (that is the question) only time can tell.

There is a rather delicious irony here: AI may make software dangerous enough that we finally have to prove that it works, and simultaneously make those proofs affordable😊.

But a caution is essential here. Anthropic did not prove these programs bug-free. The guarantees depend on the specification, execution model and trusted proof machinery. A perfectly valid proof of an incomplete specification can still leave you with perfectly valid trouble to deal with somewhere down the line.

Still, this may be one of the more consequential uses of AI that receives far less attention than chatbots.

2. Memory Is No Longer Merely Following the AI Boom. It Is Becoming Part of the Architecture

On 28 October, SK Hynix reported its Q3 results, and buried beneath the extraordinary financial numbers was something strategically more interesting.

The company says long-term agreements with 14 major customers now cover more than half of its total production capacity for 2027 and 2028. It expects memory supply to remain structurally tight through 2030, and HBM4 is expected to overtake HBM3E in both shipment volume and revenue during the current quarter.

Company forecasts deserve the usual pinch of salt though. A memory manufacturer predicting strong memory demand is not exactly an independent scientific observation, is it (I used to work for a memory company so, I know a thing or two about memory market)?

However, the contractual evidence is harder to dismiss. This means that customers are not merely talking about future AI infrastructure, but are reserving substantial future production.

More importantly, the role of memory itself is changing. The traditional mental picture was:

Processor → Memory

with the processor doing the clever work and memory feeding it.

AI increasingly makes that distinction less useful. Training and inference performance depend critically on moving enormous quantities of data between processors and memory. Bandwidth, packaging, physical proximity and energy consumed moving data increasingly determine system performance.

In other words, AI compute is becoming less about the chip and more about:

Logic + Memory + Packaging + Interconnect + Power

That is why HBM has moved from an obscure semiconductor acronym into strategic infrastructure remarkably quickly.

The uncertainty that one has to think about now is duration. Semiconductor shortages have a distinguished history of eventually becoming semiconductor oversupply. Long-term contracts reduce that risk but they do not abolish the cycle.

3. Nuclear’s Fastest Route Forward May Be through Reactors We Already Have

The energy story of the week wasn’t a futuristic reactor. It was old reactors being made slightly less old! What?

On 27 October, the US Department of Energy announced a conditional loan commitment of up to $5 billion to finance modernization and power uprates across eight Constellation nuclear plants in Illinois, Maryland and Pennsylvania. The projects are expected to add nearly 1.1 GW of generating capacity. In case you are wondering,

the projects are expected to add nearly 1.1 GW of generating capacity. In case you are wondering “What on earth am I going to do with that piece of information?”, here is something to consider: In terms of scale, what we are talking about is almost 10% of the entire nuclear generating capacity of Illinois, the largest nuclear-power state in America, where 11 reactors currently produce more than half the state’s electricity. So, this is kind of a big deal. I mean, almost 10% increase without actually building a new (and large) reactor.

An uprate is essentially an engineering effort to extract more electricity safely from an existing reactor through equipment improvements such as turbines and generators.

That matters because the electricity problem created by AI is increasingly a time problem as much as an energy problem. Why? Because new nuclear plants can take many years. And new transmission can take many years. Oh, oh, grid connections can take many years.

So, power can take many years to acquire but, data centers are being planned now. So, what is a boy to do?

Anyway, this update changes the value of existing infrastructure, given an additional gigawatt extracted from already-operating nuclear sites can potentially arrive much sooner than an entirely new generation fleet.

This does not mean nuclear has suddenly solved AI’s power problem though. You are allowed to get excited, just not too much! After all, 1.1 GW is not exactly enormous relative to projected data-center demand.

But this news suggests something subtler: When demand accelerates faster than infrastructure can be built, optimization of existing infrastructure becomes technologically important again.

We normally associate technological progress with replacing old things. But sometimes progress means discovering that the fastest new power plant is partly the one you have already built and are already using. Yup! There is still life in the old infrastructure!

4. Quantum Computing Has Found Another Possible Manufacturing Route

Quantum announcements deserve a particularly aggressive skepticism filter. Fortunately, this week’s interesting development comes with a peer-reviewed Nature paper rather than merely an optimistic press release. So, yippee!

Researchers from NVision demonstrated a single-molecule spin–photon interface using a synthetically designed organic molecule embedded in a host crystal. The work was published on 28 October.

OK, OK. I hear you. You want to know what I am talking about. Or, rather, what on earth is a single-molecule-whatever. Well, just think of it as a tiny quantum connector: The molecule can hold quantum information (useful), while the photon can carry it elsewhere (very, very useful).

Put slightly differently, scalable quantum systems need ways of connecting long-lived quantum states to photons, because photons are particularly useful for carrying quantum information between modules.

Existing approaches often rely on atomic defects in crystalline materials. Molecules offer another possibility: Chemistry allows them to be designed and manufactured from the bottom up with atomic precision. Ain’t that neat?

Still not convinced? Well, the intriguing long-term idea is not merely a better qubit (been there, not moving the needle much on its own!). It is manufacturability and scalability.

NVision is exploring whether molecular material can ultimately be deposited onto conventionally fabricated photonic chips, potentially borrowing manufacturing techniques from established semiconductor and display industries. Reuters reports that the company is working with HyperLight (a U.S.-based chipmaker) on initial devices.

That is a long way from mass-produced quantum computers. And the experiment demonstrates a promising physical interface, not an industrial process, fault-tolerant computer or commercially competitive architecture. So, be warned!

But, again, notice the recurring question appearing in quantum computing now:

Can we manufacture, connect and scale it?

This is quite a change from the previous question everyone was asking: Can the physics work?

Not sure about you but to me, this is some progress.

5. NASA Quietly Changed What “Returning to the Moon” Means in the Near Term!

On 23 October, NASA announced a significant restructuring of its human lunar-lander programme.

The agency has redefined Artemis III around a 2027 low-Earth-orbit demonstration, in which test versions of lunar landers from both SpaceX and Blue Origin would demonstrate rendezvous and docking with Orion. NASA now targets a crewed lunar-surface return in 2028, while both companies work on simplified lander architectures.

It is tempting to describe this simply as another Moon delay (well, I am a bit tempted). However, that may be missing the more interesting part.

NASA appears to be converting a difficult all-or-nothing milestone into a sequence of incrementally testable system integrations. As an engineer (and a mathematician, and physicist. Oh boy, I am confused), I am inclined to call this approach “sensible engineering”.

After all, a lunar mission isn’t one technology. Instead, it is a chain:

Launch → Spacecraft → Rendezvous → Docking → Lander → Surface operations → Ascent → Return

Testing interfaces before attempting the entire chain reduces integration risk.

There is also a structural change at play here: NASA is maintaining two competing commercial lander architectures rather than allowing the programme to depend entirely on one.

Whether that produces a 2028 Moon landing remains very much an open question. Both lander programmes still face formidable engineering challenges. So, at least for now, the correct conclusion is not about whether America will return to the Moon in 2028. That can wait for a bit longer. What is certain though is that NASA has changed the architecture by which it intends to get there.

Let’s keep watching this space…

The Connection: From Capability to Reliability

As you may have noticed in the short time I have been sharing these weekly reports with you, I am interested in connections across seemingly different stories and this week, something connects all five items.

  • AI can find bugs. Now we need AI that can prove software properties.
  • Accelerators can perform astonishing computation. Now memory movement constrains what they can actually do.
  • AI infrastructure needs electricity. Now the question is how quickly we can extract more capacity from the grid and plants already available.
  • Quantum effects work in laboratories. Now the challenge increasingly becomes manufacturing and connecting them.
  • We can design lunar landers. Now NASA wants to test the interfaces before betting an entire mission on them.

The pattern looks like this:

As technologies mature, the question shifts from “Can we make it work?” to “Can we make the whole system work reliably, repeatedly and at scale?”

That transition is easy to underestimate because reliability rarely produces spectacular demonstration videos (there is a marketing challenge here that I often wonder if someone is willing to take. I am also keen to see how demo videos of robots will change the landscape when robotics takes off soon, but we can discuss that when the time comes…).

Perhaps the most important bottleneck in advanced technology is therefore not always the one currently receiving the headlines. The main point to take from all this is that the bottleneck moves!

And when it does, value, research attention and sometimes entire industries (and, of course, investment) tend to move with it.

What I Am Watching Next

AI verification: Formal proofs generated by AI become much more interesting if independent teams reproduce the results across different architectures, operating systems and specifications. The crucial question is whether formal verification can move from specialist engineering into ordinary software development.

The memory cycle: Long-term HBM commitments are powerful evidence of real demand. The test now is whether capacity expansion catches demand faster than expected, and whether the traditional semiconductor cycle eventually reasserts itself.

Power without decade-long construction: Uprates, reactor restarts, gas turbines, storage and grid optimization are all competing against one brutal variable: Time. Watch which approaches actually deliver additional dependable megawatts rather than announcements about future megawatts.

Quantum manufacturability: Molecular spin–photon interfaces are fascinating. The next evidence required is integration, reproducibility and scaling, not another isolated laboratory demonstration.

Artemis integration: The 2027 docking demonstration should tell us considerably more about the credibility of the 2028 lunar timetable than another schedule announcement will.

Sources & Further Reading