🎧 Startups and Investment: The Week's Top Stories

_Audio podcast: two AI hosts discuss the latest AI news. Full transcript below._

Host A (00:00): You know, we've got a huge stack of fresh analysis, court leaks and financial reports on the table today. And honestly, when I was going through it all before our deep dive, it made my hair stand on end.

Host B (00:16): Yeah, there's a lot of material, and the numbers are, to put it mildly, anomalous.

Host A (00:21): Exactly, anomalous. Our main job today is to figure out how building software, just ordinary software, suddenly turned into heavy industry, a sector that needs capital on the scale of entire nations.

Host B (00:43): This shift really happened over just the last couple of quarters. The industry is no longer funding abstract code or, I don't know, algorithms in a vacuum. The current phase is an all-out, aggressive race to own critical physical infrastructure. And today we'll trace that whole chain.

Host A (01:07): Right, from completely insane AI startup valuations all the way to building real, giant fabs. The numbers we'll be dissecting today break every familiar rule of finance. I suggest we start with the software side.

Host B (01:23): Let's do it. That's where things are hottest right now.

Host A (01:25): Take the latest data on Anthropic, the team behind Claude. They're in talks to raise at least $30 billion.

Host B (01:36): 30 billion. In cash?

Host A (01:38): Yes. And the valuation tops $900 billion. But what floored me most: just 3 months ago, 1 quarter, they were valued at $20 billion.

Host B (01:49): A 45x jump in 1 quarter.

Host A (01:53): In 3 months, a garage outfit became worth more than the world's biggest banks combined. Why do they even need $30 billion right now?

Host B (02:02): Well, you have to understand the market's logic. Investors aren't pricing Anthropic's current revenue. They're essentially buying the only viable, independent alternative to OpenAI.

Host A (02:13): So they're paying a risk premium?

Host B (02:15): Exactly. But on the purely physical side, that $30 billion isn't a surplus at all. It's their survival budget for the next, say, 18 to 36 months at most.

Host A (02:27): A $30 billion survival budget. Wow!

Host B (02:31): Yes, because the money goes into building up compute. They need to buy hundreds of thousands of Nvidia H100 chips. On top of that, Anthropic says it wants to expand its model's context window to more than 200,000 tokens. Technically, in the transformer architecture, compute cost grows quadratically with context length.

Host A (02:56): Wait, let's unpack that. So if we, say, double the amount of text the model can hold in its head at once, the hardware requirements don't double, they quadruple.

Host B (03:10): Exactly right. Transformer math is relentless. Roughly speaking, every token has to attend to every other token. When you have hundreds of thousands of tokens, the number of operations becomes astronomical.

Host A (03:25): And ordinary rented servers won't cut it anymore?

Host B (03:28): 100 percent. They need their own cloud infrastructure, designed from scratch. Without this massive $30 billion injection, they physically won't be able to train the next generation. It's a war of attrition.

Host A (03:42): Wow. And interestingly, this race is rapidly turning geopolitical. In the reports I came across data on the Chinese startup Moonshot AI, which makes the Kimi chatbot.

Host B (03:55): Yeah, I read about them.

Host A (03:56): The numbers there are crazy too: they raised $2 billion at a $20 billion valuation, growing almost 7x in 16 months. But you know what's most interesting? Who's putting up the money.

Host B (04:09): The investor lineup.

Host A (04:11): Exactly. There's Meituan, the big delivery service, fine. But then there's China Mobile, a giant state-backed telecom operator, and the investment fund of the state conglomerate CATC. You have to admit, that looks nothing like classic venture capital with Valley funds at the table. It looks like direct funding of a national project.

Host B (04:34): I'd say it's a very clear signal. Beijing is literally shaping Moonshot AI into a so-called national champion. When someone on the level of China Mobile comes into the cap table, the startup gets more than a $1 billion check.

Host A (04:48): What else?

Host B (04:48): Infrastructure. Infrastructure.

Host A (04:51): And unimpeded access to national data centers, plus guaranteed public-sector contracts, and, most important for AI, access to the state's enormous internal datasets to train its algorithms.

Host B (05:05): So AI has stopped being a commercial story?

Host A (05:09): Absolutely. It's now a matter of national security and technological sovereignty at the highest level. And that, by the way, explains why the numbers are so huge. And this shift of AI into the heavy-infrastructure category brings us back to 1 very interesting court leak I wanted to talk about. It concerns Microsoft and OpenAI.

Host B (05:32): Oh yes, the hidden costs.

Host A (05:35): Right. Officially, everyone thought, and it was publicly stated, that Microsoft had invested $13 billion in OpenAI. Sounds like a lot, but manageable. Court documents have just revealed the real picture: the software giant's actual spending will exceed $100 billion by June 2026.

Host B (05:57): That's a 7x difference, mind you.

Host A (05:59): A colossal difference. It's 1 thing to train a model in a lab, and something else entirely to put it into Azure, Word and Excel and make it handle 1 million requests every second. So here's my question: if even Microsoft has to hide the real cost of running AI, does that mean generative AI has turned out to be simply too expensive?

Host B (06:24): Well, the documents show that each request to this advanced model costs tens, sometimes hundreds of times more than a regular search.

Host A (06:33): Hundreds of times?

Host B (06:34): Yes. You need huge GPU clusters running around the clock. That, incidentally, is the root of OpenAI's recent restructuring.

Host A (06:44): You mean their new spin-off?

Host B (06:45): Yes, they're setting up a separate subsidiary, a deployment company, and putting $4 billion into it right away.

Host A (06:52): Yes, yes, I saw the investor list: 19 companies chipped in for that $4 billion, including TPG, Advent International, Bank Capital, big investment players. But 1 company really caught my eye. Brookfield.

Host B (07:07): Oh, Brookfield is very telling.

Host A (07:09): One of the world's largest infrastructure and energy asset managers. They build power plants and water utilities. And they're funding this. The bottleneck isn't code anymore; it's the power outlet.

Host B (07:24): Exactly. Deploying modern AI models today is a job for an industrial developer, not a programmer. Buying servers isn't enough to build a data center. You need hundreds of hectares of land.

Host A (07:36): And water, I assume, to cool all that gear.

Host B (07:40): Yes, enormous volumes of water that have to be cleared with the authorities, and above all, a direct connection to high-voltage lines. Silicon Valley simply doesn't have that expertise. Brookfield does. This alliance is an open admission that the technology has hit the physical limits of the power grid.

Host A (07:56): Listen, since the main problems have moved into the physical world, to outlets and cooling, it makes sense to look at the hardware market. There are tectonic shifts happening there too.

Host B (08:08): And how! Just look at the Cerebras Systems IPO.

Host A (08:11): Oh yes! A $95 billion valuation. They raised more than $5.5 billion, and on the very first day of trading their shares jumped 68%. It's the largest tech listing since 2020, when Snowflake went public.

Host B (08:28): And the point isn't really the market records. What matters is why the market values them that way. They have Wafer Scale technology. In essence, they solve the main problem of modern chips: interconnect.

Host A (08:41): Meaning the physical latency as data moves between different

Host B (08:45): processors? Yes, exactly.

Host A (08:47): I came up with an analogy to make it easier to picture. Traditional chips in a supercomputer are like a bunch of small villages. For them to work on a shared task, you have to build expensive chassis and cables between them.

Host B (09:00): And run trucks full of data back and forth.

Host A (09:03): Right, and you waste tons of time and energy just on logistics. The Cerebras Wafer Scale chip is the size of an entire silicon wafer. It's like building 1 giant metropolis on 1 foundation. Everything is close together, no external cables, and the result is 40 percent more compute cores in the same area.

Host B (09:25): Great analogy. And the physics here drives the economics. Inside a single die, signals travel orders of magnitude faster and, importantly, use several times less energy than when a signal goes out to the motherboard. Training a huge neural network requires all cores to work in sync. Remove the physical distance, and you dramatically cut training time.

Host B (09:50): For giants like Meta or Google, cutting model training from several months to weeks is a critical advantage. They'll pay any price for it.

Host A (10:01): But look, Cerebras solves the speed problem. Then there's the news about Arm, which targets something else: energy efficiency.

Host B (10:12): Yes, Arm is a very important story too.

Host A (10:15): Their CEO Rene Haas said the other day that demand is exploding. In 5 weeks they got $2 billion in orders. That's 2x their norm. Arm has always been associated with smartphones, but now the growth is coming specifically from data centers.

Host B (10:33): Because data centers are fleeing classic x86 processors from Intel and AMD en masse. Why? They consume too much power.

Host A (10:42): Exactly. The x86 architecture dominated for decades, but it was designed to be general-purpose, with lots of legacy instructions. It's a CISC architecture, a complex instruction set.

Host B (10:54): Okay.

Host A (10:55): Arm uses RISC, a reduced instruction set. The chip handles only basic instructions, but executes them super efficiently. As a result, data centers save 30-40% on energy. Wow! But on the scale of 1 laptop, 40% savings is just an extra hour of TV shows.

Host A (11:14): But in a data center

Host B (11:15): Yes, in a data center with hundreds of thousands of servers, that's hundreds of megawatts. That's hundreds of millions of dollars in pure operating profit every year.

Host A (11:24): And less heat too, I assume? So less spent on cooling?

Host B (11:28): 100 percent. Less power for the servers, less power for air conditioning. A chain reaction. That's why, while the mobile market stagnates, the server segment will become Arm's main growth engine by 2026-2027. Cloud providers have no choice: they face strict limits on heat output per square

Host A (11:46): meter. Right, if you can't bring more power into the building, all that's left is to install chips that do more work per watt. Makes sense?

Host B (11:54): Exactly.

Host A (11:55): But you know, if chips have become the most valuable resource on the planet, and companies' survival depends on them, relying on outsiders becomes too risky.

Host B (12:04): Are you leading up to the SpaceX news?

Host A (12:08): Yes. It breaks every template. SpaceX is investing at least $55 billion, with potential to reach $119 billion, in an AI chip fab in Texas. The project is called TerraFab.

Host B (12:22): Insane numbers, I agree.

Host A (12:24): For context, I looked it up: NASA's entire annual budget is about $25 billion. So Elon Musk's space company plans to pour almost 5 NASA budgets into an Earth-bound chip plant.

Host B (12:38): To produce processors for 200 GW of compute capacity per year.

Host A (12:43): But why? A space company is supposed to build rockets, or satellites at most, not compete with Asian silicon fabs.

Host B (12:52): Well, you have to look at Musk's entire empire. He doesn't just have SpaceX; there's xAI, there's Tesla with its Autopilot. All these projects critically need tens of thousands of GPUs.

Host A (13:04): And he doesn't want to wait in line at Nvidia.

Host B (13:07): There's Nvidia's monopoly on design and TSMC's on manufacturing. Musk doesn't want to pay Nvidia's 70-80% margin and wait for allocations.

Host A (13:18): So it's just hard-core vertical integration. Want strong AI? Make your own silicon.

Host B (13:24): Yes. But you can't justify $119 billion with margin savings alone. Geopolitics is 100% part of this.

Host A (13:32): A shield against trade wars.

Host B (13:34): Exactly. A fab in Texas means complete independence from any blockades in the Pacific region or conflicts. It makes Musk's empire fully autonomous.

Host A (13:44): And it creates a new center of gravity in the industry, entirely outside the old supply chains. But listen, here's the most pragmatic question: we're sitting here discussing $100 billion from Microsoft, $30 billion for Anthropic, Musk's $119 billion fab. Altogether that's something like $1 trillion just being poured into AI's foundations.

Host B (14:06): Okay, so what's the

Host A (14:06): question? Who's footing the bill? Where does business actually see a return on these huge investments?

Host B (14:14): This is where the report from Bain & Company analysts helps.

Host A (14:18): Right, I read it. They estimate the future market for enterprise software built on so-called agentic AI at $100 billion.

Host B (14:28): And the key word here is agentic. We're no longer talking about chatbots that just write text or code on

Host A (14:35): request. Right, this is about automating coordination work. Sounds super boring, of course. Data entry, information exchange in ERP and CRM systems, logistics.

Host B (14:47): Sounds boring, but it costs corporations billions.

Host A (14:51): Wait, so we're really building 200-gigawatt fabs and burning the output of entire power plants just so a robot agent can fill out a boring invoice in

Host B (15:02): accounting? Yes, believe it or not. Because coordination busywork is a huge hidden tax on the economy. Millions of managers work as human routers, copying data from 1 incompatible program to another and waiting for approvals.

Host A (15:18): That's true.

Host B (15:19): Agentic AI goes into the warehouse system on its own, checks discounts in the CRM on its own, writes the report and sends it to the client on its own. It eliminates friction in business processes.

Host A (15:29): And frees people up for real work: analysis, strategy, talking to clients. In essence, it's a rebuild of all enterprise software over the next 5-10 years.

Host B (15:43): And businesses are ready to pay any amount for licenses. Every bit of time companies save will convert directly into payments for compute. It's a completely closed economic loop.

Host A (15:55): We started with context windows and valuations near $1 trillion.

Host B (15:58): And ended with concrete and power outlets.

Host A (16:01): Exactly. It turns out this seemingly ephemeral industry runs hard into physics. Plate-sized silicon wafers, Musk's giant fabs, energy-efficient cores: right before our eyes, startups have mutated into the real heavy industry of the 21st century, and Silicon Valley's rules simply no longer apply.

Host B (16:22): Yes, the sheer scale of capital has become the main filter. Without connections in global energy and industrial construction, there's nothing left for you in AI.

Host A (16:30): Listen, since we're on energy, I had 1 thought at the end, a provocative one, that wasn't directly in these reports.

Host B (16:38): Okay, interesting. Go on.

Host A (16:40): Look. If SpaceX needs 200 GW, and Arm is fighting to the death for every 10% of watt savings in data centers, what if over the next 10 years the main bottleneck isn't a shortage of chips or investor money at all? What then? A plain lack of capacity in the world's power grids. There simply won't be enough throughput.

Host A (17:04): Won't tomorrow's tech giants have to become the world's largest energy companies? Build their own nuclear plants just so their algorithms don't go dark for lack of power?

Host B (17:15): That's a very strong question. And judging by deals like the one with Brookfield, they've already started.

Host A (17:20): Exactly. Let's leave that thought to mull over. That wraps up our deep dive. Until our next plunge into the data.