Tag: Biotechnology

Weekly Technology Brief #1

From Bigger Models to Bigger Systems

8 October 2026

Dear readers,

Here we go, the very first weekly technology brief…

If there is one idea running through this week’s technology news, it is that the interesting action is moving outward. AI is no longer just a story about better models (this has, arguably, been the case for the best part of 2026 so far). The constraints, opportunities and risks are increasingly appearing in the systems around them. This includes the agents that act on our behalf, the memory chips that enable lots of the fancy work, the electricity that keeps the data centres running, the biological datasets that let the models reason about living systems, and the verification machinery needed before we trust claims about quantum computers, autonomous robots, or the incredible number of mathematical proofs thrown at the scientific community.

That shift matters because it changes the question. Instead of asking only, “Which model is smartest?”, we increasingly need to ask: What infrastructure makes the intelligence useful? What becomes the next bottleneck? And, perhaps most importantly, how do we know that the system actually works outside a carefully controlled demonstration (this last one itself deserves a deep dive soon)?

1. The Enterprise AI Race Is Becoming a Control Problem

Google Cloud used its Gemini at Work event this week to introduce what it calls a single universal Gemini agent for work. The important part is not that the agent can answer questions, create content or write code. After all, we have seen plenty of that already. The more interesting part is where Google is putting the surrounding machinery.

The agent is designed to operate across Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, while carrying business context between them. Google is also emphasizing identity, permissions, authorization, secure sandboxing, network gateways, policy management and spending controls. In other words, the product is beginning to look less like a clever chatbot and more like an employee operating inside an institutional control system. At least that is the idea. Whether or not any of the individual angles are well covered is a matter of further verification and a source of more articles!

However, this is an important transition. A model that occasionally produces a bad answer is irritating. An agent with permission to send an email, alter a spreadsheet, query confidential data, execute code or initiate a business process can create a very different class of problem. As the ability to generate actions becomes cheaper, the value of deciding which actions are permitted, and independently checking what actually happened, should rise. Also, let’s not forget that verifying, and ultimately pulling the plug on, a model generating a single wrong answer is easy. Once an agent gets to work, the potential for generating nonsense also scales and it becomes quite possible that by the time we know what is wrong the extent of the any damage cause could be truly substantial.

In any case, the transition means that the emerging stack may actually look something like this:

Model → Agent → Permissions/Control Plane → Execution → Verification.

The model companies will naturally compete on intelligence. Enterprises, however, may eventually care just as much about the layers around that intelligence, including identity, audit trails, rollback, cost control, data sovereignty and liability.

As a result of this transition, intelligence itself may become abundant, but trustworthy control probably may not necessarily follow the same trajectory.

2. AI’s Electricity Problem Is Turning Existing Nuclear Plants into Technology Infrastructure

The physical side of AI became unusually visible this week when Google announced a major agreement with Constellation Energy. The arrangement is intended to unlock 890 megawatts of additional nuclear capacity through upgrades to 11 reactor units across six existing U.S. nuclear plants. Google says its agreements involving uprates and reactor restarts have now enabled more than 1.5 gigawatts of new nuclear capacity.

The interesting detail is that this is not primarily a story about waiting for futuristic small modular reactors. Google is financing improvements to infrastructure that already exists. That is revealing. When electricity demand is rising quickly, the fastest new megawatt may be the one squeezed out of an existing asset rather than a completely new power station.

This is another example of the bottleneck migrating downstream. Better AI models create demand for more accelerators. More accelerators create demand for memory and networking. More servers create demand for data centers. Data centers create demand for power, cooling, land, grid connections, permits and financing.

Build this, and that becomes the next bottleneck!

The implication is broader than nuclear power. AI infrastructure is becoming an industrial system, and industrial systems are constrained by things that software companies historically did not have to worry about quite so much, such as turbines, transformers, transmission lines, water, construction schedules, local communities and the cost of capital. The future of AI may still be written partly in code but an increasing amount of it will be built in concrete and copper. Sounds strangely poetic, right? 😊.

3. Memory Is Still Telling Us Where the AI Bottleneck Lives

Samsung’s preliminary third-quarter results provided another reminder that the AI boom is not confined to GPU vendors. The company projected operating profit of 107.4 trillion won for the quarter, nearly nine times the level a year earlier (12.17 trillion), with strong demand and tight supply in memory playing a central role. Detailed results are still to come later this month, so the precise divisional picture deserves some patience.

The more useful lesson is structural. Modern AI accelerators can perform extraordinary numbers of calculations, but those calculations are useful only if data can be moved to and from the processors quickly enough. That is why high-bandwidth memory has become such an important part of the AI system, and why conventional DRAM and NAND conditions still matter as capacity is redirected and supply remains tight.

There is also a warning here. Extraordinary earnings do not prove that extraordinary growth continues indefinitely. Semiconductor markets have a long history of turning shortages into investment, investment into capacity and capacity into eventual oversupply. AI may have changed the scale of demand, but it has not repealed economics. In other words, the inherent cyclicality of memory products remains a real risk for investors.

So, the question is not simply whether memory demand is strong. It plainly is. The better question is whether supply can expand faster than AI workloads broaden. And where the constraint moves if memory becomes less scarce. Networking? Advanced packaging? Power? Capital? This is why following bottlenecks is often more informative than following product announcements.

We used to say (well, I used to have this as my LinkedIn moto for a long time): In data we trust. As an investor, especially in the crazy age of AI one should say: In bottlenecks we trust!

4. AI for Biology Is Discovering That Data Must Sometimes Be Manufactured

One of the most interesting developments of the week came from biology rather than the usual AI laboratories. The U.S. government, Biohub, Meta, Google DeepMind and Isomorphic Labs are participating in a roughly $1.8 billion effort to build large, standardized datasets for AI-driven biological research.

This matters because biology exposes one of the weaknesses in the familiar internet-era AI story (let me be clear: Weakness in this context is not exactly a negative word. If you get biological data wrong and then use it to make a decision that may subsequently prove wrong, people may die. I mean, literally. That is why anything to do with biology has been traditionally more regulated and, by definition, slower. And yes, that includes data availability. While there is always room for improvement everywhere, including in biology, expecting every industry to run under the start-up maxims of “break it first and see where it ends”, or “fail fast” is not exactly advisable!). Language models benefited from enormous quantities of text that already existed. In biology, the most valuable training data may not exist yet. Researchers have to create it through experiments. This is not my area of expertise but I read this includes perturbing cells, measuring what happens, standardizing the observations, then training models, generating new hypotheses and finally running more experiments.

In other words, the loop becomes:

Question → Experiment → Data → Model → Prediction → New/Refined Question → New/Refined Experiment → …

Each cycle uses the evidence generated by the previous one to refine the next question and experiment.

That is potentially much more consequential than simply attaching an AI assistant to a laboratory (there is nothing wrong with the AI assistant but, eh!). If models can help decide which experiments are most informative, and automated laboratories can execute those experiments at scale, scientific discovery itself becomes a closed learning loop. Granted, scientific discovery was always expected to be loopy (using feedback to refine questions and experiments) so, the concept may not be new. What is new is our ability (thanks to modern AI) to accelerate and increasingly automate parts of that loop, and to do it at scale.

But there is an important discipline to maintain here. A better biological model is not a medicine. Neither is a plausible mechanism the same as a clinical outcome. Drug development still has to cross the brutal sequence from biological insight to candidate, preclinical evidence, human safety, efficacy and eventually scalable treatment (I just told you about the regulations and the slower pace of change in biology, did I not?). AI may accelerate parts of that chain without magically eliminating the chain (I am, assuming, that biology will not fall into a Muskian loophole any time soon. Remember DOGE?).

5. Quantum Computing Is Entering the ‘Show Me’ Phase

Quantum computing produces no shortage of impressive announcements (on a personal note, and as a Physics graduate, I have to admit that there is a lot to be excited about in this field these days. It is so, very, grooooovy!). This week’s more interesting development was deliberately less glamorous though: DARPA moved Atom Computing, Diraq, IBM and IonQ into Stage C of its Quantum Benchmarking Initiative. They join Microsoft and PsiQuantum, which had already reached the equivalent final evaluation stage.

DARPA’s question is refreshingly concrete: Can any of these approaches produce a utility-scale quantum computer whose computational value exceeds its cost by 2033 (as I said in my separate intro before publishing this very first weekly brief, my aim is NOT to give my readers investment advice. I am simply not an expert in that. But the investor in you may want to pay attention to the clear deadline introduced here. What might credible progress, or failure to make it, mean for company valuations between now and 2033? Now, that is an interesting question.)? Stage C moves the emphasis from roadmaps toward verification and validation of the hardware and systems as they are built.

That is exactly the kind of pressure the field needs. Qubit counts, fidelity records and elegant demonstrations are useful evidence, but none alone tells us whether a machine can solve economically valuable problems reliably enough, at sufficient scale and at acceptable total cost.

The deeper pattern is worth watching beyond quantum:

Claim → Execution → Measurement → Independent Verification → Trust

As technology becomes more powerful and the financial incentives to announce breakthroughs become larger, verification itself becomes valuable infrastructure (who will do the verification? Does this smell like a possible business opportunity, I wonder!).

6. Robotics Is Beginning to Build Its Own Safety Layer

A recent, smaller announcement this week may be more important than many flashy humanoid demonstrations. Agility Robotics and FORT Robotics expanded their work on a safety architecture for Digit 5, including interfaces that connect the robot to external safety systems.

That sounds less exciting than a humanoid folding laundry or performing acrobatics (and even less exciting than a video of a humanoid folding laundry while performing acrobatics!). But I say that is good. In fact, bring more of this type of news on! I learned a long time ago that real deployment is supposed to become boring. Else, the said deployment may not be nearly as reliable as one would hope.

A robot operating beside people in a warehouse or factory needs more than an impressive policy model. It needs predictable emergency behavior, communications with external (or 3rd party) safety infrastructure, clear operational boundaries and ways for humans or other systems to intervene. The relevant scoreboard therefore should not be “How human-like did the demo look?” but something closer to useful task completion, autonomous operating hours, intervention rate, safety incidents and cost per useful hour.

Physical AI will become interesting when robots stop being primarily demonstrations of capability and become dependable pieces of infrastructure.

The Connection: Intelligence Is Moving into the Real World

Taken separately, these stories concern enterprise software, nuclear energy, memory chips, biology, quantum computing and robotics. Taken together, they point in the same direction.

We spent the first phase of the current AI boom staring at models. The next phase is forcing us to look at systems:

  • Agents that need permissions and verification,
  • Accelerators which need memory,
  • Data centers that need electricity,
  • Biological models which need experimentally generated data,
  • Quantum roadmaps that need independent validation,
  • Robots which need safety infrastructure.

This is why I increasingly find the word “AI” too small for the technological change we are trying to understand. No, I am not going to suggest Super Intelligence. I mean, give me a break!

The model may be the intelligence engine, but the economic and social consequences emerge from the larger system built around it.

What I Am Watching Next

Three questions stand out for me. First, whether enterprise agents genuinely move from answering questions to completing useful work without creating an unacceptable control and verification burden. Second, whether the AI infrastructure bottleneck keeps migrating from chips and memory toward electricity, grid capacity and financing. Third, whether AI-for-science systems begin producing experimentally validated discoveries faster (and not merely more convincing predictions).

Those are harder tests than benchmark scores or product demonstrations. They are also, I suspect, much closer to where the real value will be created.

The question for next week’s brief remains the same: What happened in the prior 7 days that meaningfully changed our understanding of where technology is going?

Sources & Further Reading