🎧 Startups and Investments: The Week's Main Stories

_Audio podcast — two AI hosts discuss fresh AI news. Full transcript below._

Host A (00:00): You know, we have on the table today just a massive pile of fresh analytics, all sorts of court leaks and financial reports. And honestly, when I was reviewing all this before our deep dive, my hair stood on end.

Host B (00:16): Yeah, quite a bit of material has piled up, and the figures are, to put it mildly, anomalous.

Host A (00:21): Exactly, anomalous. And our main task today is to try to understand how software development, ordinary software, suddenly turned into heavy industry—a sector that requires capital on the level of entire nations.

Host B (00:43): This paradigm shift actually happened literally over the last couple of quarters. We see that the industry is no longer financing just abstract code or algorithms in a vacuum. The current phase is this, you know, maximally aggressive race to control critical physical infrastructure. And today we'll analyze this entire chain.

Host A (01:07): Yeah, from absolutely insane valuations of AI startups to building real giant factories. The figures we'll be dissecting today genuinely break any conventional financial logic. And I propose we start with the software sector itself.

Host B (01:23): Let's. It's the hottest thing right now.

Host A (01:25): Well, for example, fresh data on Anthropic. These guys created the Claude neural network. They're currently negotiating to attract at least $30 billion.

Host B (01:36): 30 billion. Cash?

Host A (01:38): Yes. And the company valuation exceeds $900 billion. But what got me the most is that just 3 months ago, in Q1, they were valued at $20 billion.

Host B (01:49): A 45x growth in one quarter.

Host A (01:53): A garage became worth more than the world's largest banks combined in 3 months. Why do they even need 30 billion dollars right now?

Host B (02:02): Well, look, you have to understand market logic. Investors are not evaluating 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 if you look at the purely physical side of things, this $30 billion is not excess. It's their survival budget for the next 18 to 36 months maximum.

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

Host B (02:31): Yes, because the money goes to acquiring computational capacity. They need to buy hundreds of thousands of NVIDIA H100 chips. Plus Anthropic claims it wants to expand its model's context window to more than 200,000 tokens. And from a technical standpoint, in transformer architecture, computational complexity grows quadratically relative to context length.

Host A (02:56): Wait, let me clarify for understanding. So if we increase the amount of text the neural network can hold in its head at once by 2x, hardware requirements don't grow by 2x but immediately by 4x?

Host B (03:10): Absolutely correct. Transformer mathematics is unforgiving. Each token must, roughly speaking, attend to every other token. When you have hundreds of thousands of tokens, the number of mathematical operations becomes just astronomical.

Host A (03:25): And you can't get away with regular rented servers?

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

Host A (03:42): Yeah. And this race, interestingly, is rapidly acquiring geopolitical scale. I came across data in reports about Chinese startup MoonshotAI, which makes the Kimi chatbot.

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

Host A (03:56): They also attracted crazy figures—$2 billion at a $20 billion valuation, growing nearly 7x in 16 months. But, you know, what's most interesting is who's giving the money?

Host B (04:09): The composition of investors.

Host A (04:11): Exactly. There's Meituan, a large delivery service, okay. But then China Mobile—a giant telecom operator with state participation—and an investment fund of the state conglomerate CATC. But you have to admit, this looks nothing like classic venture, where funds from the Valley are sitting around. This looks like direct financing of a national project.

Host B (04:34): I'd say this is a very clear signal. Beijing is essentially molding MoonshotAI into what's called a national champion. When someone at China Mobile's level enters the capital, the startup doesn't just get a check.

Host A (04:48): What else?

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

Host A (04:51): And unimpeded access to national datacenters, plus guaranteed contracts in the government sector, and most importantly for AI, access to colossal masses of state internal data for training their algorithms.

Host B (05:05): So artificial intelligence 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 this, by the way, perfectly explains why the figures are so enormous. Yeah, and this transition of AI into the category of heavy infrastructure brings us back to a very interesting court leak I wanted to discuss. It concerns Microsoft and OpenAI.

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

Host A (05:35): Yes. Officially, everyone thought it was publicly stated that Microsoft invested $13 billion in OpenAI. That sounds like a lot, but tolerable. But court documents just revealed the real picture. The software giant's actual costs will exceed $100 billion by June 2026.

Host B (05:57): A 7x difference, mind you.

Host A (05:59): A colossal difference. It's one thing to train a model in a lab and quite another to shove it into the cloud, into Word, into Excel and make it process a million requests every second. Here's the question I have: if even Microsoft is forced to hide the real costs of maintaining AI, doesn't that mean generative AI turned out to be just too expensive a luxury?

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

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

Host B (06:34): Yes. You need round-the-clock operation of huge clusters of graphics processors. That's where OpenAI's recent restructuring finds its roots, by the way.

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

Host B (06:45): Yeah, they're creating a separate subsidiary and immediately throwing $4 billion into it.

Host A (06:52): Yes yes yes, I saw this investor list—19 companies chipped in for these $4 billion, including TPG, Advent International, Bank Capital, big investment guys. But one company really caught my attention. Brookfield.

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

Host A (07:09): The world's largest infrastructure and energy asset manager. Meaning they build power plants, water systems. And here they're giving money because—the bottleneck became not code. The bottleneck became the outlet.

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

Host A (07:36): And water, probably, for cooling all this stuff.

Host B (07:40): Yeah, colossal amounts of water that need to be coordinated with authorities, and most importantly direct connection to high-voltage lines. Silicon Valley simply doesn't have this expertise. But Brookfield does. This alliance is an open acknowledgment that technology has hit the physical limit of power grids.

Host A (07:56): Listen, since the main problems have shifted to 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! What Cerebras Systems did.

Host A (08:11): Oh yeah! A $95 billion valuation. They attracted more than $5.5 billion, and on the first day of trading their shares jumped 68%. And this is the largest tech IPO since 2020, when Snowflake went public.

Host B (08:28): And here, you know, the point isn't even in exchange records. What matters is why the market values them so highly. They have Wafer Scale technology. They essentially solved the main problem of modern chips—the interconnect problem.

Host A (08:41): That is, physical latency, when data runs between different...

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

Host A (08:47): I came up with an analogy to make it easier to visualize. Traditional chips in a supercomputer are like a bunch of small villages. And for them to solve a common task, expensive chassis and cables need to be built between them.

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

Host A (09:03): Yeah, yeah, tons of time and energy wasted just on logistics. And Cerebras' Wafer Scale chip is the size of an entire silicon wafer. It's like building one giant megacity on one foundation. Everything close, no external cables, and as a result 40 percent more computational cores for the same area.

Host B (09:25): Great analogy. And the physics of the process determines the economics. Inside one crystal, signals fly orders of magnitude faster, and what's important, require far less energy than when a signal leaves for the motherboard. Training a huge neural network requires all cores to work synchronously. By removing physical distance, you dramatically shorten 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 amount for that.

Host A (10:01): But look, Cerebras solves the speed problem. But there's also this news about Arm, which hits a different target—energy efficiency itself.

Host B (10:12): Yeah, the Arm news is also very important.

Host A (10:15): Their CEO Rene Haas recently announced an explosion in demand. In 5 weeks they received $2 billion in orders. That's 2x higher than their norm. And Arm was always associated with smartphones, but now this growth comes from datacenters.

Host B (10:33): Because datacenters are massively fleeing from classic x86 processors from Intel and AMD. Why? They consume too much energy.

Host A (10:42): Exactly. The x86 architecture dominated for decades, but it was created as universal, with tons of legacy instructions. It's a CISC architecture, complex instruction set.

Host B (10:54): Right.

Host A (10:55): And ARM uses RISC, reduced instruction set. That is, the chip does only basic instructions, but does them super efficiently. As a result, datacenters save 30-40% energy. Wow! Listen, but on a laptop scale, 40% savings is just an extra hour of watching shows.

Host A (11:14): But in a datacenter...

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

Host A (11:24): And less heat, probably? That is, less spending on cooling?

Host B (11:28): 100 percent. Less electricity for servers, less electricity for air conditioners. Chain reaction. That's why the mobile market is stagnating now, and the server segment will become ARM's main growth engine toward 2026-2027. Cloud providers simply have no other choice; they have strict limits on heat output per square meter.

Host A (11:46): Yeah, if you can't bring more power to the building, the only option 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 became the most valuable resource on the planet, on which the survival of companies depends, relying on someone else becomes too risky.

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

Host A (12:08): Yes. This just breaks all patterns. SpaceX is investing at minimum $55 billion with potential up to $119 billion in an AI chipmaking factory in Texas. The project is called TerraFab.

Host B (12:22): The figures are insane, I agree.

Host A (12:24): For context, I looked up NASA's entire annual budget—about $25 billion. So Elon Musk's space company is about to drop nearly 5 NASA budgets into an earthbound microchip factory.

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

Host A (12:43): So why? A space company should build rockets, maybe in extreme cases satellites, but not compete with Asian factories for silicon smelting?

Host B (12:52): Well, you have to look at Musk's whole empire. He doesn't just have SpaceX; there's xAI, there's Tesla with autopilots. 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 monopoly on production. Musk doesn't want to pay NVIDIA's 70-80% margin and wait for quotas.

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

Host B (13:24): Yeah. But you can't justify $119 billion in margin savings alone. There's definitely geopolitics here.

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

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

Host A (13:44): And creates a new gravity center in the industry altogether, outside the old supply chains. But listen, I have the most practical question here—we're sitting here discussing $100 billion from Microsoft, $30 billion from Anthropic, Musk's factory for $119 billion. That's already pushing toward a trillion dollars just being poured into AI's foundation.

Host B (14:06): So what's...

Host A (14:06): the question? Who pays for the banquet? Where does business actually see a return on these gigantic investments?

Host B (14:14): And that's where a report from Bainin Company analysts helps.

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

Host B (14:28): And the key word there is agentic. We're not talking about chatbots that just write text or code on...

Host A (14:35): request. Yeah, we're talking about automation, coordination work. Sounds, of course, super boring. Data entry, information exchange in ERP and CRM systems, logistics.

Host B (14:47): Sounds boring, but it's worth a billion dollars to the corporation.

Host A (14:51): So wait, we're really building 200 gigawatt factories and burning electricity from entire power plants just so a robot agent can fill in a boring invoice form in...

Host B (15:02): accounting? Yes, imagine. Because coordination routine is a huge hidden tax on the economy. A million managers work like living routers, copying data from one incompatible program to another, waiting for approvals.

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

Host B (15:19): Agentic AI itself climbs into the warehouse system, checks discounts in the CRM, writes the report and sends it to the client. This eliminates friction in business processes.

Host A (15:29): And frees up people for real work, analysis, strategy, customer communication. Essentially, this is a complete rebuild of all corporate software over the next 5-10 years.

Host B (15:43): And business is ready to pay any amount for licenses. Every hour companies save will be converted directly into payment for computational capacity. This is a completely closed economic model.

Host A (15:55): From context windows and valuations in the hundreds of billions...

Host B (15:58): to concrete and outlets.

Host A (16:01): Exactly. Turned out that all this seemingly ephemeral industry is rigidly bound to physics. Silicon wafers the size of plates, giant Musk factories, energy-efficient cores—startups mutated before our eyes into real heavy industry of the 21st century. The rules of Silicon Valley just don't work anymore.

Host B (16:22): Yeah, capital scale became the main filter. Without connections in global energy and industrial construction, there's nothing left to catch in AI.

Host A (16:30): Listen, and since we're already talking about energy, at the end a rather provocative thought occurred to me, one that wasn't directly in these reports.

Host B (16:38): Interesting. Let's hear it.

Host A (16:40): Look. If SpaceX needs 200 GW, if ARM is fighting for its life for every 10% savings in Watts in datacenters, what if in the next 10 years the main bottleneck won't be chip shortages or investor money deficits? What? Plain capacity shortage in the world's power grids. Just not enough bandwidth.

Host A (17:04): Will tomorrow's tech giants be forced to become the world's largest energy companies? Building their own nuclear power plants just to keep their algorithms from dying from power shortage?

Host B (17:15): That's a very strong question. And judging by the deals with Brookfield, they're already starting to do it.

Host A (17:20): Exactly. Let's leave that thought for reflection. And that concludes our deep dive. Until the next immersion in the data.