The Year in Review - the year in Infrastructure

It's been a crazy year for digital infrastructure with trillions of dollars promises for new AI data centers, associated bubbles, the dawn of agents, and significant progress in adjacent standards, tools and processes.

Massive growth in planned AI infrastructure has driven massive growth in AI stocks. The bubble almost burst, but didn’t – yet. Meanwhile, MCP for Agentic AI integration has sparked developer enthusiasm not seen since the early days of mashups and Web 2.0. Investors, regulators and Big Tech are all looking for ways to grow AI like the early Internet, which turned out not to be a sure thing.

A bevy of new standards, techniques and a promising alternative to neural networks for the spatial web and reality capture are helping close the gap between AI and the physical world. Also, there is not enough power and water for those new data centers, but maybe there is, or at least could be. Innovators are also starting to reimagine tricky processes for construction compliance from the ground up.

Jeffrey Emanuel and the lessons we should all learn from the $2 trillion DeepSeek AI market correction

I think a lot of people are very smart, high IQ, they understand semiconductors, they understand AI, but then they try to start talking about valuing a stock, and I'm like ‘Oh this is a completely different game.’ If you don't know what you're talking about, it's very easy to make a complete fool of yourself.

Why? Jeffrey Emanuel’s essay on popular misconceptions in AI infrastructure was widely credited with wiping out over $2.7 trillion in market capitalization and $600 billion from NVIDIA’s market cap alone in a short period. It seemed like a real coup to interview and get a sense of how that particular essay came to be and the nuance he brought to his analysis.

Many of the structural foundations of NVIDIA’s moat still remain. Yet cloud providers are developing alternative chips, and companies like DeepSeek are demonstrating that it's possible to train highly capable models at lower cost and give them away for free. It's also a cautionary tale of calling the bubble too soon. Emanuel is also bullish on AI's long-term prospects.

Don't panic! What Sam Altman’s OTT declaration of a 'code red' for OpenAI tells us about the race he thinks he's running

What will really cause Nvidia to cut price and take the gross margin hit will be direct competition for hardware, either from Google selling TPUs externally (but even if that happens, they’ll surely still make a good gross margin, since why wouldn’t they?) or from new entrants making custom inference silicon. But it could take a while.”

Why? That’s from the same Jeffrey Emanuel ten months later in December, after NVIDIA’s stock had recently raced from $3 to $5 trillion since the previous interview. Oh, and Sam Altman declared a ‘code red’ after Google delivered a better LLM that quickly took on 200 million new users. In the short run, this is probably a good thing for users since OpenAI is likely to focus more effort on building a better model rather than creative enshittification productization strategies.

This and the preceding story paint a more nuanced picture of what may or may not be an AI bubble today. Some really savvy investors predicted the Internet bubble four years before it burst, but missed out on tremendous opportunities. And despite numerous questions about OpenAI’s circular financing, money-losing products, and setbacks in relative LLM performance, it's not necessarily game over for them. Emanuel observes they still have a massive consumer brand, a loyal installed base and will be monetizing it more going forward.

Agentic mashups - just like getting the band back together...

It's almost like the bands are back together in a funny way, where so many of the people who were in that period are struck by and engaged by the MCP AI era. This whole notion of what's going on today in 2025 has so many similarities to what was going on twenty years ago. It's a bit of a cultural moment, if you will. The same but different.

Why? Former ProgrammableWeb Founder and CEO John Musser had a front-row seat from the API revolution's inception to its maturity, which utterly reshaped how apps were built and connected. This phenomenon was variously called mashups, Web 2.0, API governance, and eventually just the way apps were built. In the process, it cleaned the slate of legacy and complicated Service Oriented Architecture and Enterprise Service Buses, which took a long time to die a good and proper death.

These days, he is seeing the same pattern repeat with the birth of Model Context Protocol (MCP) as a new paradigm for connecting AI agents, tools, and enterprise apps. Many of the same players from the first era are jumping back into the game with newfound enthusiasm and vigor to reimagine enterprise processes beyond the limits of one-trick chatbots.

Lessons in system engineering the Internet and why it matters today

There are many theories about why TCP/IP was more successful than its contemporaries, and they are not readily testable. Most likely, there were many factors that played into the success of the Internet protocols. But I rate congestion control as one of the key factors that enabled the Internet to progress from moderate to global scale. It is also an interesting study in how the particular architectural choices made in the 1970s proved themselves over the subsequent decades.

Why? The Internet infrastructure we have today came as a bit of a surprise to the major telco and cable operators, which had quasi-monopolies on digital circuits in the early 1990s, charging usurious fees. And it's probably a good thing for all of us today that they chose not to participate widely in its architecture or regulations. Certainly, in its early days, the TCP/IP network was full of bugs and limitations, and even Ethernet inventor Bob Metcalfe cautioned it would flame out any day.

But it didn’t. Systems engineers like Bruce Davie, Larry Peterson, and many others worked together to strengthen the foundation supporting the multi-trillion-dollar digital economy we have today. The lesson today is that, as much as people point to limits in scaling laws, LLM hallucinations, money-losing companies, and data centers no one needs, smart minds can come together to work around these limitations.

How the Spatial Web could guide de-centralized and trustworthy agentic AI to like us more

Space and time help us organize information. We’ve got the most general version of space and time in hyperspace. What could we do with that? We could do collective intelligence. The notion with collective intelligence is that we will have an ecosystem of intelligent agents interacting more efficiently and in new ways, because we have a modeling language, a transaction protocol, and an agent framework that does not currently exist.  The Spatial Web will enable this collective intelligence to perform activities and meet humanity's challenges that we couldn't do before.

Why: IEEE Spatial Web Vice-Chair George Percivall explains how the new spatial web will open opportunities to connect autonomous physical systems with networks of AI agents. It's not just about sharing data, as with existing APIs and even MCP approaches. It represents a paradigm shift with primitives for data governance, managing uncertainty, identity, certifications, smart contracts, and knowledge graphs in an extensible way.

It's exciting and early days. NASA has demonstrated a proof of concept of agentic robots collaborating on the moon. But it will take a few years to test the independent implementations, define precise specifications without ambiguity, and test interoperability at scale. Percivall, who was previously CTO of the Open Geospatial Consortium for seventeen years, observes that these sorts of things often seem to move faster in the future than expected.

Growing better brains - why we need to re-think the neuron for more trustworthy and efficient AI

Currently, with all Large Language Models, you've got large in the title. This is what it says on the tin. They start off overly expressive, overly parameterized, overly funded, and then reduce it down. You can imagine fine-tuning a foundation model as part of this journey to get back to a simpler, more specialized architecture.  But now, in principle, we have another way of doing it, which is much more efficient, which is to approach from below, but that requires you to be able to grow models.

Why: One of the biggest limitations of LLMs is that they only capture relationships in the words people use to describe the world, rather than causal models of the world. Verses Chief Scientist Karl Friston is a rare breed of scientist who bridges animal brain research and AI, with over 2,000 papers to his name. He believes that Bayesian networks that mirror human brain processes are a better fit for causal models that treat uncertainty estimation as a first-class primitive.

Early work on an approach called active inference is already showing impressive progress in agentic AI systems compared to the reinforcement learning approaches popular today. That said, it's early days. More work will likely be required to build infrastructure to scale collections of these agentic brains, share knowledge, and align them with LLMs and other techniques for distilling human knowledge.

Why crossing the pixel/reality gap is essential for better digital twins and UI

A glaring gap in the existing crop of smart glasses and VR/XR headsets on the mass market today is the mismatch between how our eyes perceive the world (3D light fields) and how they render content (2D stereoscopic images). Surprisingly, there has been little discussion on this issue by vendors or the mainstream press. This will change when recent innovations in portable light field tech start to hit the mass market in 2-3 years, and more people start to appreciate the difference.

Why: Lightfield technology for capturing and rendering 3D fields the way our eyes see the world is not exactly new. The first light field camera was demonstrated by Nobel Prize winner Gabriel Lippmann in 1908, and the field took another seminal leap with the discovery of holography in the 1960s. But the equipment for capturing, storing, processing, and rendering these has always been expensive and inflexible.

Things took a major step forward this year with numerous innovations in Gaussian splats and Neural Radiance Fields for lightfield-based reality capture and highly realistic 3D rendering. Over the last year, these advances have been finding their way into Meta’s new teleportation app, construction digital twins, and supporting real-time bulk inventory monitoring in supply chains. What’s more, NeRFs and Gaussian Splats are a natural fit for AI and neural network infrastructure. Just around the corner lies the commercialization of lightweight lightfield (or holographic) kit for rendering more realistic and ergonomic smart glass displays.

Data centers and AI do use water - but less than you think, and there are much worse offenders

The total global operational data center capacity today is estimated at 42.4 GW. In other words, the water lost by utilities in the UK every day could almost cover cooling demand for the entire global data center IT footprint, which highlights where attention should really be focused.

Why: Power and water infrastructure costs are being widely raised as essential constraints for all those data centers Big Tech wants to build. Gary Flood has taken a critical look at this narrative across this and other coverage throughout the year. Water limitations may indeed be an issue in some places across the US, with reports of water quality issues and price hikes. Data center operators, particularly those building the largest ones, are also pioneering the adoption of new closed-loop cooling processes that dramatically reduce water constraints and improve power usage efficiency.

In the UK, water issues attributed to data centers end up being a much smaller issue compared to leaking pipes and fat executive payouts taken against investment in water infrastructure. As I write this, 20,000 people in Kent, UK have been without potable water for the past week. New AI data centers seem like an easy target for this rage, but water problems in the UK have been an issue long before the data center straw man took center stage.

How Europe is re-wiring its power grid with open source digital twins

It’s hard to separate when you want to know what's going to happen. There's a physical level of a grid, then there is a software level, then there is an economic level, and all that goes together.

Why: In contrast with water, electricity is a major constraint to all the planned AI data centers, not to mention electric cars, heat pumps, and more electrified industry. It's not entirely clear how operators will build the new AI data centers, given 2-5-year queues to bring on new power plants of all types in the US, UK, and Europe. This is a complex problem spanning existing market paradigms, legacy wires, and investment challenges.

Linux Foundation Energy has made recent progress on two simulation tools to improve analysis and collaboration among finance, infrastructure planning, and operations teams. One surprising insight was how cultural differences, just within Europe, shape all these domains. Coopetition to develop the best open source components to shape the operation of a shared grid across the EU serves as a form of soft power.

How Spacial is re-imagining construction compliance processes with AI

Automating code compliance is hard because building codes aren’t necessarily digital-friendly. They are legal documents written in ambiguous, jurisdiction-specific language. Every city, county, or state might enforce a slightly different interpretation of similar rules. Standards like IFC and newer BIM interoperability have helped improve model structure, but they don’t solve the semantic complexity or local nuance of interpreting real-world code enforcement.

Why: There has been progress in digital building information models and construction regulation standards developed by engineers. It's easy to imagine that automating the permitting process would be a straightforward problem to solve. There is the adjacent problem of designing buildings that are easy to construct and include all of the structural, mechanical, and electrical systems that keep occupants safe, happy, and energy use down.

Spacial has taken a step back from just rushing out an AI tool and reimagined the process, keeping engineers in-the-loop with AI development and fine-tuning. It's time-consuming. They have to move market by market and work with local permitting authorities to interpret ambiguous legal text and incorporate their intentions and cautions into the process. But this is helping them align the output with city review expectations, using traceable logic and reviewer-friendly formats.