£102bn Failure, AI Bills, and the Construction Robots Shipping Now

This week, the official or should I say the current official verdict came in on the most expensive piece of infrastructure ever attempted in the UK. The numbers are absolutely eye-watering. Maybe eye-watering is a slight under-adjaculation here. Tens of billions of pounds over budget. A decade late, at least a decade late. Two of the most respected reviewers in British engineering pulled apart what went wrong and they landed on the same root causes. And as far as I can tell, nobody named in any of it will lose their job over what happened. That story sits at the centre of what we are going to talk about today because it's really about whether the UK construction industry can still deliver anything at the price the rest of the world pays for it. While that was paying out, something else happened. The cost of AI that every construction software company is now built on top of just went up and you wouldn't even have noticed. Very quiet and very sly by the AI companies. The pricing pages don't show it. And unless you dig really deep into these things you'd never even realise. Finally, the robots the industry has been waiting nearly two decades for accelerating deployment on real job sites and generating real revenue for contractors. So let's get into it. Before we do though, a quick announcement from this week's sponsor, Bradcrumb. One poorly managed incident can demolish your project's entire profit. Bradcrumb replaces the paperwork, safety plans, orientations, permits, all digital in any language, seeing straight into ProCore. Over one million workers and 50,000 plus projects trusted. Visit breadcrumb.co details in the show notes. Right. Let me start with the AI cost story. This is really hiding in plain sight. If you run any kind of technology budget inside of a construction business, you need to understand what's happening. If you've been listening to my briefing for a while, you remember a couple of months ago I told you that the cost of intelligence had dropped 280 times in around two years. That number came from a senior construction leader I spoke to and at the time it felt like the most optimistic data point in the industry. The price of AI was a little bit more than the cost of AI. The friction was no longer the cost of AI. It was how you organized your people around it. Two months on, that picture is kind of changing because the headline is still true. Per token prices have collapsed. Some of the cheapest models on the market today cost less than 30 per cent per million tokens. That's about the price of running about 350,000 words for an AI model for roughly the cost of a coffee. But here's what's also true and which nobody else is really noticing except for the people on the receiving end of it. Enterprise AI bills are exploded. Some companies are now seeing monthly AI invoices in the tens of millions of dollars. And in the same week or maybe two weeks or so, there's two biggest AI labs in the world both raise their prices in ways that nobody noticed. Let me explain how. So open AI a few weeks ago launched a new version of their flagship model, 5.5. It was 5.4 recently and doubled the rate on the pricing page. They told everyone the new model was less talkative so the total cost would balance out. However, independent data from the largest model marketplace in the world looked at users who switched from the old version to the new version, so from 5.4 to 5.5 and ran actual numbers. All costs went up between 49 and 92%. In the same week, Anthropic released a new version of their flagship model, 0.4.7. I think this was around about a month ago and they didn't change the rate card at all. What they changed was something called a tokenizer. So let me just explain that. A token is a chunk of text. The model uses the count what is reading. That's the meter. And Thropic's new meter counts the same English sentence as up to 35% more tokens. So the price per token is identical but the same prompt now costs more. The pricing page of course does not show that. And these are two different mechanisms with the same outcome. But the headline the price went down but the actual bill people are experiencing is going up. This to me feels like another one of those cheeky tricks that AI companies use to either restrict your usage or somehow get you paying more for something that you now so heavily rely on. And by all accounts this problem will continue to accelerate and only get worse for us users. Now here's why it matters for people running construction businesses. Almost every new construction product being pitched or in fact every new construction technology product being pitched to you right now is built on what's called an agentic workflow. You don't need to know the engineering or what's underneath the hood. What you need to know is that a regular chat bot asks the AI one question and gets one answer. An agentic workflow asks the AI 10 or 20 questions complete one task because it's doing things like reading documents, checking its own work, calling other tools and then verifying the output. So when prices per token go down by 50%, the volume of tokens use per task goes up by 10 to 20 times. The maths is brutal. You pay less per unit but you consume far more units. Your total bill goes up. This is what economics call the G1's paradox. When something becomes more efficient, you don't use less of it. You actually use more because all new kinds of applications that didn't make sense at the old price suddenly do at the new one. The classic example is steam engines. When they got more efficient at burning coal, total coal consumption went up, not down. The exact same thing is happening with AI right now and it's happening underneath every single AI product that you are being sold. But here maybe is the good news for you. So this week I had a long conversation with a founder who saw this coming maybe two years ago. She runs a company building natural language software for construction litigation and she's been doing it for about 10 years. When I spoke to her, she told me something that stuck with me. Back in 2024, when she was talking to the big AI labs, they would all tell her the same thing. Don't worry about it. The cost of running AI will go down. Compute will get cheaper. Models will get smaller. Throw all your data at the model and let the model figure it out. She thought they were lying. Not deliberately, just commercially incentivized to make builders more comfortable with the consumption patterns they were going to come back and buy them. So what she's saying was that these companies want you to get comfortable with using a model in a certain way so that when they turn the dial on the pricing, you're stuck in your old habits and old ways of doing things and so they can ultimately charge you more money. So she built her company the opposite way. Her system takes 100% of a client's data and pre-sorts it down to about 1% of what actually matters. Then, and only then does she let these expensive and ever growing more expensive AI models touch that 1%. Her phrase was, "You don't need to kill a rabbit with a bazooka." So next time, instead of throwing a thousand pages as an AI model and running a single-line prompt which will cost you an absolute bomb, her company reduces it down to 10 pages that actually matter so when you turn your prompt an AI output is 99% cheaper. In a world where AI is getting more expensive, that is a game changer. She also said something that stuck with me. She said, "The big trend in the next six months and this is her prediction is going to be unlocking legacy data at scale." Whereas contractors have decades of project records, emails, claims files, drawings, and so on, all sitting in formats that the new AI models can't easily use. The companies that figure out how to decode that legacy data without lighting their AI budget on fire are the ones who win the next phase. That logic, by the way, is the exact opposite of what some senior figures in industry have been telling contractors recently. There's been a school of fault seated by Azat Brickson Bytes in our discussion with Martin Fisher from Stanford that says, "Throw the old data away. It is a liability and the future is built on new, clean data only." Based on what this founder is saying, "I'm not so sure anymore. Legacy data sitting inside your business might be a moat for you and you probably shouldn't completely disregard it as having no use." Let me bring another voice into this conversation because recently I sat down with an innovation leader at Shimizu, who are a large mega general contractor in Japan. They're a 220 year old company and I have about 20,000 employees that listed on the Japanese stock exchange they are a huge business. The innovation leader, Jean-Marc, was previously running corporate venture for British telecom in Silicon Valley for over 20 years. So he's seen the AI hype cycle from inside the labs that created it and he's now applying that inside of a major contractor, A.E. Shimizu. He told me something specific. He said, "Shimizu rolled out a general AI tool to half their 20,000 staff in just seven months. The tool they chose isn't from open AI or Microsoft directly. It's from a Japanese startup called Light Blue. A Tokyo university spin out that Shimizu had previously invested through their corporate venture fund. What Light Blue gave them was a single secure front door into all the major AI models, plus the ability to point those models at Shimizu's own project documentation. 5,000 of his colleagues onboarded in the first wave, which is the rate of adoption that almost no construction business in Europe or America has matched. But here's a bit I want you to really pay attention to. When I asked Jean Mark about what he thinks about the cost of all of this AI tokens, he said the following. Right now we're in an AI subsidy. Everything is cheaper than what it really should be. We don't see the real cost of the AI. And the way that these models are priced is going to change. His warning to anyone running an innovation function was this. If you can't show a hard benefit for what you're spending, your management is going to come down on you like a ton of bricks and shut the whole thing down. is a man. running AI inside a major contractor in 2026 telling you in plain English that the bill is coming. And when it does, the firms who can't prove ROI will be the ones canceling their AI programs first. Think about it like this. As AI becomes more critical to the operation of a business and even the entire industry, the companies who figure out how to reduce their AI spend while still getting optimal outputs will naturally have a cost competitive advantage over any of their competitors. And now is the time to act on that particular knowledge. Just to reverse slightly and go back to what light blue does. And I think it's important to highlight this because this will be a growing area within the world of AI. What they do is rather than you having to put your prompt into a tool like Claude, Chat GPT, Gemini, whichever one you use, you put it into light blue and it then chooses the model for you based on what you're asking and what your output is and then it will give you the most optimal foundational model to basis response offer there by potentially saving you cost and giving you the best output for what you are putting in to the machine. And just like I said about the founder earlier, who was building something kind of similar and you'll almost certainly see this area of AI, let's say, optimization growing over the next months or years as the models do gradually get more expensive. So what do you do with all of this knowledge? What's your finance lead and your IT lead, the same question separately and see if their answers match? The question is what did our AI spend look like in April compared to March and what specifically are we getting for it? If they can't answer with a number and my guess is that they can't you have a problem that will only get worse every single day from here. I'd love to hear how your business is tracking AI spend right now. Drop me a comment on the LinkedIn post for this episode and tell me, are you measuring it? Have you seen the bill? Do you know what your team is using and what it's costing? Let me know in the comments. Let's move on to the second story of this week. Robotics. I know we have talked about robotics a lot on this podcast over the last few weeks and I don't want to be too idealistic, but the wave is coming. People are getting a lot more interest in the field now and there's much more technology being developed to make it possible on construction sites, which you'll get to in a second and a lot more conversations and interest from people who are thinking three to five years ahead within their organization. So I hope on this podcast we can keep you up to date and give you some insight as to what is happening in the field of robotics. So at Brickson Bytes, we published a piece of research called The State of Physical Artificial Intelligence and Robotics in 2026. I'll put a link of this in the show notes and the whole report runs across the six biggest industries that physical AI is moving into. Now physical AI is a word that you will hear much more often. I'm sure you've already seen it. Now I want to talk you through the part that matters within that report for the construction industry. So after 60 years of construction being the last sector to be automated, the first real wave of construction robots has now graduated from demos to repeat deployments with real contractors generating verified return on investment. The reason that sentence is now true is down to five structural barriers that have been holding the industry back, each of which is finally starting to bend. Let's go through them quickly. But before we do a few weeks back, we covered the humanoid on a construction site. It's exciting news, but what I'm going to be talking about in this section is a much more reflective discussion on where the industry is at right now and a more reflective discussion on the broader industry. Not everyone has humanoids on construction sites, for example. Okay, the five structural barriers. The first one is that every construction project is unique. A factory, for example, builds the same component a million times. Construction site builds something that has never existed before, often from a design that still changing the morning of the concrete poor old style robots needed repetition to be economical. The new generation can work from building information model updated that morning. So the barrier of uniqueness is starting to bend for tasks where the individual actions are repetitive, even if the overall project is one of a kind. The second is the genuinely messy environment of a building site uneven ground, whether by a half-billion structures, multiple trades working at the same time, working from high and so on. Modern robot vision systems can now tolerate dust and clutter, where the previous generation couldn't. Not every task on a life site has cleared the bar yet, but enough of them have to make a real commercial difference. The third is industry fragmentation. A typical commercial project has a main contractor, 20 or 40 subcontractors, a design team, owner and multiple suppliers, nobody can force technology adoption on anyone else. So the robots that win are the ones a single trade contractor can deploy without needing buy in from anyone else in the project. That's a specific filter that rules out most of the splashy demos that you see at the trade shows. The fourth is thin margins. When you run on two to five percent margin and sometimes even less, you cannot justify a 500,000 piece, $500,000 piece of hardware up front. That's why price models at charge per square meter or per pile installed or per scan are the only ones that work typically in this industry. Pure hardware sales kill more construction robotics companies than any technology failure does. The fifth is the sequential nature of the work. Every trade depends on one on the one before. If the trade in front of you is late, the robot you deployed sits idle. So the robots that succeed are the ones deployed early in the sequence where the dependency risk is lowest. Earth moving, layout, foundation work, re-biteying, the boring stuff at the start of the job. So with all of that in mind, here's what's actually shipping right now. The biggest category by far is autonomous, heavy equipment, excavators, bulldozers, soil compactors, all running on software stacks that came directly out of the self driving car industry like Waymo, layout on top of machines from caterpillar, commatsu, Volvo and John Deer. And the kit can be installed in one to three hours and removed just as quickly. Construction sites are fenced. They're repeatable and they don't have the long tail of edge cases that made self driving cars so hard for so long. A startup called Bedrock Robotics raised $270 million and now valued at over $1 billion. I think the total that they've now raised is in fact $350 million. And another company called Crue Line has raised $7 million. It's just a first, four person team in April and immediately had a $26 million order book waiting for them. These are not pilots. They are revenue generating production deployments on real American job sites. The second category is re-biteying. Re-bite is heavy, repetitive, dangerous work that sits right on the critical path of every concrete poor. Human re-bite workers tie 40 to 80 intersections an hour. The robot ties 300 to 450. On a big industrial map with 200,000 intersections, that compresses 12 days of work into 4 to 6. The third category is autonomous piling for solar farms and ground works. The world installed over 380 gigawatts of solar last year. Most of it sits on piles driven into the ground by robotic rigs that retrofit onto existing equipment. The economics here are genuinely transformative. The fourth category is reality capture and digital inspection. The scanning of a site that feeds into a structure project record. This is becoming a standard expectation on large commercial projects, not a premium add-on. And then there's a fifth category which is just emerging as probably the most interesting one. A single robot that completes a structural element and immediately verifies its alignment, fastener talk and dimensional tolerance, then file the record back to the project recorded automatically. The labor saving is real but the bigger prize is the data trail. Every project produces structured, comparable building wide records that make the next project faster, cheaper and lower risk to finance. Now I know what some of you are maybe thinking because there's a few weeks back we talked about the humanoid robotics on this briefing and you've seen all the humanoid robotics doing back flips and folding laundry and so on. So where are they in construction? The only answer based on most ground in research I've read is that large scale humanoid use in construction is probably a decade or so away. That's not me being pessimistic or a report being pessimistic as a specific quantitative case. For mass deployment and construction humanoid units cost need to fall between 20 to between 20 and 50 thousand dollars per humanoid and it needs to be competitive with the local labor rates which can vary enormously as you know by region. China hits that crossover point earlier than Europe or the US. The need to have a task for humanoid in construction are narrow and structured. It can do site inspections, reality capture, maybe even unloading trucks. Meaning out specific places are preparing tools for skilled trades. The point of humanoid in this decade is to remove low skilled overhead so that skilled trades people can concentrate on the work only humans can do. Anyone that tells you that humanoid are about to replace your brick layers on Monday morning is probably trying to sell you something. So the companies that are actually winning right now are the ones building purpose built task robots which is exactly where most of the investment capital is flowing. Construction technology robotics pulled in what? $1.36 billion in the first nine months of 2025 more than double the entire previous year. That money is being priced very differently depending on the workflow. Heavy equipment autonomy and reality capture are being funded at around $100 million per round on average. Interior and finishing robots where the technology is less mature are funded at around $27 million per round. That four times gap is the market selling you which workflows it believes in and which it's still watching. But here's the part of the research I think every construction person needs to internalize. The single biggest reason construction robot its companies fail is not technology it's the business model. The candid version from founders and investors who've watched it happen goes like this. Contractor's willingness to take on unproven technology is essentially zero. That's not a negotiation position at all. It's a structural reality of a thin margin, liability-driven industry, where a failed technology deployment costs not just the hardware, but the schedule impact, the claims exposure, and the reputation damaged with a client who hired you. The burden of proof for unproven construction robotics is much higher than in manufacturing or logistics. The first commercially available bricklaying robot launched in 2015 at $500,000 per unit. Almost nobody bought one. The company's been very public about how hard that journey was, and the lesson is that if you come to a contractor with a robot promising a whole new workflow, and you leave all the risk with them, nobody is going to buy. The only models that work are subscription, leasing or pay per outcome, per pile installed, per square meter laid, per scan completed, and so on. There's a great line from one of the founders of Canvas, which is the most successful drywall finishing robot its company in the industry period. The co-founder Maria actually got her own drywall finishing license and physically worked alongside the robot on early deployments. When she explained why, she said the workers on site knew there was a human with manual labor skills there, if anything went wrong with the robot. That made it much easier for them to trust the technology, physical presence and operational credibility, not just technology demos is how trust gets built in construction. So here's where I turn to you. Of all the work players that I mentioned, what do you think is going to show the first repeat purchase pattern at scale this year? Will it be heavy equipment autonomy, rebar tying, solar piling, and so on? Or one of the emerging ones like finishing off a side work? Let me know what you think in the comments. [Music] Let me come back to where I started because I owe the rest of that crazy story about the infrastructure project. The infrastructure that it is at the top of the show is HS2. The transport secretary, Heidi Alexander, confirmed in Parliament this week that the cost of complete is now between £873 billion. The first services between London and Birmingham will not open until somewhere between 2036 and 2039. The original promise back in 2012 was that the full network to Birmingham, Manchester and Leeds would be operational this year in 2026 at cost of £32 billion. Let me say that again. £32 billion, not £102.7 like they are now predicting. That is the most expensive piece of infrastructure ever attempted in the UK. And the price tag is frankly a national embarrassment when you compare it to the cost of the world. HS2 is costing more than £400 million per kilometre. France builds its high speed rail for about £20 million per kilometre. Spain does it for about 15. Britain is paying somewhere between 15 and 20 times the international rate for the exact same technology. Three separate reviews went into what went wrong. And they all blame the same things that the specification was gold-plated. The trains were designed to go faster than any train in the whole wide world which made every component be spoke. The main civil's contracts were led to international joint ventures before the designs were even ready which transferred risk those contractors couldn't even price. The commercial model was target cost on paper but turned into cost plus in reality, which rewarded contractors regardless of how badly the project ran. The client organisation HS2 Limited was the liberally set up as a lean team, which meant it lacked the commercial muscle to manage what it had bought, and the sponsor, the department for transport, did not scrutinise. The board packed data, according to the most recent review, was, and I quote, "a veil behind which less good news becomes difficult to assess" whatever that means. The man now running it, the running this reset in fact, is Mark Wilde, who delivered the cross rail recovery. He told the Parliamentary Committee this week that the cost overrun in his words is mostly in efficiency of work because we started too soon. Designs weren't ready, contracts were let, anyway the bill kept growing because the work kept agreeing done badly, and the contracts had no mechanism to stop it. Now here's why I'm telling you all of this and why I led the episode with it. You can spend the next week arguing about whose fault HS2 is, the Conservatives who started it, the civil servers who didn't scrutinise it, the joint ventures who priced it badly, the politicians who kept cutting the scope. All of those arguments have some merit, and none of them change what happens next because the same people will be allowed to bid on the next mega project, and the same model will probably be used to deliver it. Here's some fun stats for you. So the total cost is £102.7 billion for a train track. The cost in 2011 was predicted to be £32.7 in 2013, £45 billion, in 2015, £55.7 billion, 2019, up to £88 billion, 2020, up to £106 billion, 2026 up to £102.7 billion cheaper, but a heavily reduced route. The original plan was eight cities, over £340 miles, 32 billion pounds, and open. In 2026, and what we're getting, four stations, £140 miles of track down from £340, and at a cost of over £100 billion open in 2039, and guess who is paying you and me the average taxpayer. Okay, enough ranting, and let's try and wrap this week up as I realize I have gone over our time allowed for this episode. So on the 3rd of June, 2026, it's digital construction week in London, and we will also be hosting our own small event alongside Lang O'Rook. And I will be having a fireside chat with Chetan Kotour. Chetan, he was a car designer at Volvo, who joined the founding team that spun Polestar out as a global electric vehicle brand. He scaled that business from 100 people to roughly 4,000, known to say in 29 markets and built factories across three continents. Then, Raya Rook, the founder of Lang O'Rook, made him a phone call, and today, Chetan leads a 50 person team at Lang O'Rook, trying to re-engineering how a tier one main contracted designs, manufactures and builds. This is the exact kind of conversation that we love to have on here. Somebody actually trying to rebuild the main contractor model from the inside and the person doing it, not coming from construction at all. We'll also be joined by the Suffolk Tech team, who are flying in from Boston, and we've already had a very strong response from multiple senior executives across the industry. The room is deliberately capped, but if you're going to be in London for digital construction week and you need a construction business, then drop me a message and I'll put a link in the show notes for the episode. So before I let you go, three things that you can do this week. Firstly, ask your finance leader and your head of IT separately while your AI spend looked like. Secondly, start to immerse yourself in the world of robotics. It is real. It's coming. Everyone you speak to is super excited about it, or be it a little bit skeptical, but the wind is really starting to change and it could be one of the most important things you focus on over the next few years. So, thirdly, don't be like HS2. Learn from the lessons that we have just shared on the show. And don't go spending all of our taxpayer's money. If this briefing was useful for you, please share it and drop me a comment on the LinkedIn post for the episode. We'd love to hear from you. And on that note, I will see you all next week.