Tag: AI technology

Prince Mario-Max Schaumburg-Lippe: Nvidia-Backed Firmus Plans $5.5B IPO at $30.6B Value

The Biggest AI Infrastructure IPO of the Year

The AI boom has a new kind of landmark deal. Australian data center operator Firmus plans an initial public offering of up to $5.5 billion, at a share price that values the company at about $30.6 billion. Reuters reported the details on October 5, citing a term sheet and people familiar with the matter.

If it lands, this will be the second-largest Australian-listed IPO on record, behind only Telstra’s $10 billion share sale in 1997. The bookbuild begins Tuesday, October 6, with the institutional close pulled forward to Thursday because investor indications have already come in well above the offering size. Trading on the Australian Securities Exchange is expected to start October 23, with the prospectus lodged October 12.

Bank of America, JPMorgan, Morgan Stanley and Morgans are leading the deal. Firmus itself declined to comment.

Who Gets the Shares

Here is the detail that tells you how hot this is: roughly half of the IPO, including the over-allotment option, is earmarked for existing strategic and financial investors. The company’s disclosed backers include Nvidia, Coatue, Blackstone and Jane Street. According to reporting on the term sheet, Nvidia holds about 7.2 percent, Coatue around 8.4 percent, and Blackstone roughly 6.7 percent, and the allocation lets them top up at the listing price rather than watch their stakes dilute.

The valuation math is dizzying. Firmus raised a $2 billion strategic equity round in August, with Nvidia and Coatue making follow-on investments and Blackstone and Jane Street participating. That round valued the company at about $10.5 billion post-money. The IPO price of A$11 per share implies a valuation of about $30.6 billion. Nearly tripled in two months.

Some of the money is already spoken for in the physical world. Proceeds are earmarked for GPUs at the company’s first data center in Batam, Indonesia, part of a previously announced plan to deploy 170,000 Nvidia GPUs at the campus.

What Firmus Actually Builds

Firmus is not a software company. It does not train models or sell subscriptions. It builds the physical layer the models run on: modular AI data center platforms, designed for high-density compute, with proprietary cooling and power engineering.

The footprint tells the story. The company has a presence in Singapore, one of Southeast Asia’s primary interconnection hubs, and a facility in Melbourne that demonstrates its platform can scale inside Australia’s enterprise and government digital ecosystems. The IPO proceeds will fund a global rollout of the modular platforms, multi-gigawatt grid interconnections, expanded manufacturing for cooling modules, and next-generation high-bandwidth hardware.

That list is worth reading closely. Land, power rights, cooling innovation, specialized modular design: these are the bottlenecks of the AI era. Chips get the headlines, but a GPU without power, cooling, and a building around it is a paperweight. The companies that control the physical layer are emerging as the critical chokepoints of the whole supply chain. Nvidia’s 7.2 percent stake is the industry’s way of admitting it: the chipmaker needs world-class places to put its silicon, and it is buying into the companies that build them.

The Infrastructure Supercycle

The Firmus listing is arriving in the middle of a historic capital wave. Goldman Sachs just raised its year-end 2026 US data center capacity forecast to 64 gigawatts, and analysts estimate US power demand from data centers will grow 38 percent this year. A Bain analysis projects annual AI infrastructure spending could reach $1.5 trillion by 2031, which would require the industry to generate around $6 trillion in yearly revenue to justify it.

Those are the kinds of numbers that make a $30.6 billion valuation look like the beginning of a cycle, not the end of one. Every model launch, every agent platform, every robotics round like FieldAI’s $700 million raise ultimately cashes out in megawatts. Someone has to build the buildings.

There is tension in the story, and it is worth naming honestly. Data centers face growing public opposition over electricity demand and local impacts; only a fraction of Americans say they would welcome one in their community. Firmus’s modular, efficiency-focused approach is partly an answer to that: better cooling and higher density mean more compute per megawatt, which is the metric that matters to grids and neighbors alike.

The Takeaway

A $5.5 billion IPO for a company that builds buildings for computers sounds absurd until you remember what those buildings do. Every frontier model trains and runs inside exactly this kind of infrastructure, and the self-hosted trend IBM pushed this week only adds to the demand: the more companies want AI running in their own buildings, the more buildings need building.

The Firmus listing is the market putting a price on the pick-and-shovel layer of the AI gold rush. Thirty billion dollars, tripled in two months, with demand already above supply. The next few years will test whether the revenue can catch up with the concrete. But the direction is not in doubt: AI runs on power, and power runs on companies like this one.

If you want to know where AI goes next, watch the power contracts and the cooling patents, not just the benchmark charts. The $5.5 billion number is the headline. The multi-gigawatt grid interconnections are the story.

Prince Mario-Max Schaumburg-Lippe: OneByZero Raises $20M to Take AI Into Enterprise Asia

The Hardest Step in Enterprise AI

Every large company now has an AI pilot program. Most of them also have a graveyard of AI pilots that never went anywhere.

That gap, between a promising demo and software running inside real workflows, is the hardest problem in enterprise AI. It is also the business that Singapore’s OneByZero just raised $20 million to solve.

The company announced on October 5 that it closed a $20 million Series A led by Jungle Ventures. It is OneByZero’s first external financing. The money will fund expansion across Asia Pacific, a new local team in Japan, and continued development of the company’s NEO platform and its AI agents.

What OneByZero Actually Does

OneByZero is not a model company. It does not train foundation models and it does not sell a chatbot subscription. It helps large enterprises integrate AI into the systems and workflows they already run: the finance department’s reconciliation process, the telecom’s customer operations, the retailer’s supply chain planning.

The company works with large enterprises in finance, telecommunications, and retail, and says it now operates across nine markets in Asia Pacific and the United States. The details of the customer list are not public, but the operating model is the interesting part. OneByZero puts engineers close to customers, working inside their workflows, rather than shipping generic software from a distance.

That forward-deployed model has a real trade-off. Engineers embedded with customers can build deeper integrations and higher switching costs, and they accumulate knowledge about how an industry’s work actually gets done. But it is more labor-intensive than selling pure SaaS, and it scales at the speed of hiring. The question investors are asking is the same one the whole enterprise AI market is asking: does close-in integration compound into something defensible, or does it just sell hours?

Why Asia, and Why Now

The geographic bet is deliberate. Asia Pacific’s largest enterprises are sitting on enormous operational complexity: multi-country supply chains, dense regulatory regimes, workforces that mix languages and systems. AI adoption there has lagged the US narrative, but the demand is real, and the companies that crack deployment in these environments build playbooks that are hard to copy.

Japan is the tell. OneByZero is building a local team there, which suggests the company has learned what every enterprise AI vendor eventually learns: in Japan, you do not sell software from a Singapore office. You show up.

The timing lines up with a broader shift. The enterprise conversation has moved from “which model is smartest” to “which vendor can get it into production.” IBM’s self-hosted coding platform made the same bet this week from the infrastructure side: the winner is whoever handles the unglamorous parts, security reviews, data residency, integration with the ancient system nobody wants to touch.

The Series A Market in 2026

Twenty million dollars is not a headline number in 2026, and that is part of the point. The mega-rounds get the press. FieldAI is reportedly raising $700 million. But the Series A tier is where the AI economy is actually being built: dozens of companies like OneByZero, raising real money to do the deployment work the labs cannot do.

Jungle Ventures is betting that OneByZero has crossed one of enterprise AI’s harder barriers: moving customers from experiments into production. In a market full of companies selling potential, a company selling working deployments at nine markets’ scale is a different animal.

The NEO platform and the AI agents the company is building deserve a watch. If OneByZero can productize what its engineers learn inside customer workflows, the labor-intensive model becomes an asset instead of a cost. Every deployment makes the next one faster. That is the flywheel the whole services-meets-software category is chasing.

The Takeaway

The AI industry loves to talk about intelligence. The money is increasingly flowing to something less glamorous: integration.

OneByZero’s $20 million raise is a bet that the bottleneck is not smarter models but better deployment, and that Asia Pacific’s enterprises will pay well for someone who does the hard part. It is the same lesson the agent infrastructure wave is teaching on the cloud side and agentic recruiting is teaching in HR. Pilots are cheap. Production is the product.

Watch the Japan expansion. If OneByZero plants a real team there and it works, the playbook is proven. And the graveyard of enterprise AI pilots gets one more resident rescued.

What Enterprises Should Actually Do

There is a practical lesson here for any company still stuck in pilot mode. The vendors worth betting on in 2026 are the ones who talk about your existing systems first and their models second. Ask them how they handle your data residency rules. Ask them who shows up when the integration breaks at 2 a.m. Ask them to name three customers in your industry who made it to production, and what broke along the way.

OneByZero is not the only company selling this promise, and Jungle Ventures’ check does not guarantee it delivers. But the thesis it represents is the healthiest one in enterprise AI right now: intelligence is abundant, and deployment is the scarce skill. The companies that master the scarce skill win the decade.

Prince Mario-Max Schaumburg-Lippe: Brett Adcock’s Hark Launches This Week: AI for Everything

The Robot Guy Wants to Run Your To-Do List

Brett Adcock has spent the last few years building humanoid robots. Now he wants to handle your dinner reservations.

On October 4, the Figure founder and CEO posted on X that Hark, his personal AI company, will launch this week. The offer is aggressive: the first 100,000 registered users get the paid plan free. A waitlist is open now.

If you have not heard of Hark, you are not behind. The company operated in stealth for months and only surfaced publicly in late 2025, when Adcock revealed he had put $100 million of his own money into the project. Since then, it has grown into one of the most ambitious bets in consumer AI: a personal AI that remembers your preferences, works across the websites you use every day, and eventually connects to dedicated hardware built just for it.

What Hark Actually Is

Forget the chatbot comparison. Hark’s pitch is an AI that does things, not one that answers questions.

The clearest preview of that vision came on August 5, when the company showed off Hark Handoff, a research preview of its browser agent. Handoff drives a virtual computer: it opens a browser, clicks, scrolls, types, reads files, and runs terminal commands. The work it is aimed at is refreshingly ordinary: placing food orders, shopping online, booking restaurant tables, researching and arranging travel.

That last point is the design choice that matters. Handoff interacts with websites the way people do, clicking through real pages, instead of depending on each service to build a separate integration. That means it works with sites that have no public API. The trade-off is honest too: it also means Hark depends on websites that can change their layouts, block automated activity, or demand human verification checks. The company will be fighting that battle on every site it touches.

The longer-term vision is bigger. Hark’s manifesto describes a system that builds a rich, evolving understanding of its user, keeps persistent memory across conversations and tasks, and eventually connects to dedicated hardware built just for it.

The Money and the Team

Hark has funded this ambition at startup-superstar scale. The company has raised more than $700 million in Series A capital, and it assembled a team of 45 engineers and designers early on, including former Meta AI researchers and designers from Apple and Tesla. There is also a strategic thread running through Adcock’s empire: Hark’s models are already being trained on data from Figure’s robots, and the company secured a deal with Nvidia for thousands of GPUs for training.

Adcock will keep running Figure as CEO alongside Hark. The two companies are separate, with no announced plan to merge, but the overlap is obvious: robots that understand the physical world and personal AI that understands your life are two halves of the same idea.

Adcock says he now uses the product for everything. He did not say how long the free paid plan lasts for those first 100,000 users, or what exactly it includes. Details like that usually surface at launch.

Why Launch Week Matters

The consumer AI agent space has been all promise and very little product. Every demo video shows a flawless agent booking the perfect trip. Almost none of them survive contact with real websites, real edge cases, real CAPTCHAs.

That is exactly why a real launch matters. DigitalOcean spent last week packaging agent infrastructure into one monthly bill, because agents are getting serious enough that the machinery around them is a business. Metaview raised $60 million to put agents to work in recruiting. The agent economy is moving from slides to products. Hark is the first big bet that the consumer side can work too.

The 100,000-user free offer is the classic consumer playbook: remove every reason not to try it. Adcock is betting that once people hand their errands to an agent that remembers them, they will not go back to doing it themselves. He is probably right about the psychology. The question is whether the product is ready.

The Takeaway

Hark is either the start of the post-app era or a very expensive lesson in how hard the real web is. Both outcomes are interesting.

If you are one of the curious, the waitlist is open and the first 100,000 paid plans are free. If you are one of the skeptical, fair: a research preview in August is a long way from an agent you can trust with your credit card. The honest move is the same for both groups. Watch this week’s launch for one thing only: does it handle the boring stuff, reliably, on the websites people actually use?

That is the whole test. Agents that can answer hiring questions or move boxes in warehouses are already proving themselves in narrow lanes. Hark is trying the wide lane: everything, for everyone. Nobody has pulled that off yet. This week, we find out if the robot guy is the one who does.

Prince Mario-Max Schaumburg-Lippe: Aleph Alpha Launches Kolibri Sovereign AI Model

A Hummingbird Lands on German Reunification Day

The timing was deliberate. On October 3, 2026, the Day of German Reunification, Aleph Alpha released Kolibri. Kolibri is German for hummingbird, and the name fits the engineering: a 78.1-billion-parameter model that only ever uses about 3.5 billion of them at a time.

The release landed on the company’s blog under a headline that made no attempt at subtlety: “Kolibri Has Landed: A Sovereign Open-Weight Model.” The full weights went up on Hugging Face the same morning, under the Apache 2.0 license. That last detail matters more than the poetry. Apache 2.0 means anyone can download Kolibri, run it, fine-tune it, and ship commercial products on top of it, without asking permission or filling out a form.

This is Aleph Alpha’s answer to the biggest question in European AI: can Europe build a serious model of its own, on its own terms?

The Engineering in Plain Terms

Kolibri is a mixture-of-experts model. Think of it as a model built from hundreds of specialists instead of one generalist. In each of its 50 layers, a router picks 6 experts out of 384 to handle the current token. Total parameters: 78.1 billion. Active on any given token: roughly 3.46 billion, or about 4.4 percent.

The point of that split is cost. All 78 billion parameters have to live in memory (about 78 GB in FP8, so the minimum hardware is two NVIDIA A100 80GB GPUs), but only a fraction of them burn compute on each token. You get the knowledge capacity of a giant model with the running cost of a much smaller one.

The context window is long: 262,144 tokens natively, extendable to a million through configuration. The model ships with four reasoning effort levels (none, low, medium, high) and tool calling. It handles English and German, with German making up about 21.3 percent of pre-training data, English around 62 percent, and code about 14 percent. Aleph Alpha even built a custom tokenizer, UniBPE, with a 128,000-token vocabulary tuned for German compound words, so German text costs fewer tokens to process.

Training itself is part of the pitch. Kolibri was trained on 768 NVIDIA B200 GPUs in Germany and Finland, under European and German law, on roughly 24 trillion tokens. Before committing to the full run, the team validated the pipeline on a smaller sibling, Kolibri Origin (30 billion total, 3 billion active, 65k context), then scaled the same approach up. Pre-training stayed stable across hardware failures and dropped connections without human intervention, which at this scale is not a small achievement. One faulty node at 768 GPUs usually means a dead run and a 3 a.m. pager.

The Benchmarks, Honestly Framed

Aleph Alpha publishes its numbers, as every lab does, and they should be read the way all vendor benchmarks are read: as a starting point, not a verdict.

Kolibri scores 96.9 on AIME 2025, 84.3 on GPQA Diamond, and 85.9 on LiveCodeBench v6, per the company’s reporting. The more interesting chart is not a ranking. It plots average score against decoded text per second per GPU, against Kolibri Origin, Qwen3.6-35B-A3B, Nemotron 3 Super, and Mistral Small 4. Aleph Alpha claims Kolibri sits on the Pareto frontier there: best quality for the serving cost, against models with up to four times its active parameter count.

Two details stand out. First, the company ran its math and science benchmarks in German and published that column alongside the English one. Almost nobody does this, and it is exactly what a model pitched at German public administration should be doing. Second, the model is signed to the EU General-Purpose AI code of practice, which is the compliance story European customers actually need to hear.

The model card also notes Aleph Alpha designed Kolibri to refrain from answering when it lacks supporting evidence, an anti-hallucination stance that matters for mission-critical use. Take it as a design goal to verify in practice, not a solved problem.

Sovereignty You Can Download

“Sovereign AI” gets thrown around a lot. Kolibri gives it a concrete meaning. Aleph Alpha uses the word in two senses: how the model was built (by teams in Germany, on infrastructure in Germany and Finland, under European and German law) and how it reaches customers (open weights they can run in their own data centers).

That second part is the real story. It is the same direction the enterprise market has been moving all month. IBM made its coding agent platform self-hostable this week, letting companies keep code and context inside their own walls. Kolibri takes the idea further down the stack: the model itself, downloadable, Apache-licensed, yours to run. No API key. No vendor with a kill switch. No sensitive documents traveling to someone else’s cloud.

“Kolibri demonstrates that we have the talent and expertise in Germany to develop competitive AI models,” said CEO Ilhan Scheer. “For us, AI sovereignty means freedom of choice by retaining the ability to build and advance this technology, and giving customers control over how they use it.”

Why It Matters Beyond Germany

The open-model conversation in 2026 has been dominated by the US and China. Reflection is reportedly about to ship the American answer. DeepSeek and Qwen set the bar the American models are chasing. Kolibri makes the map triangular: a European open-weight model, built under European law, benchmarked in German as well as English, and released under a license that lets businesses actually use it.

For regulated sectors, hospitals, banks, aerospace contractors, and government agencies across Europe, the calculation is simple. The best closed models are brilliant and unusable for your most sensitive data. Kolibri is the attempt to close that gap: frontier-adjacent quality, two GPUs of hardware, your building, your rules.

It will not be the biggest model of 2026. That is not the point. Kolibri is proof that a 200-person team in Heidelberg can ship a serious open-weight model on its own infrastructure, in months rather than years, and hand it to the world under Apache 2.0. The hummingbird landed. Watch what it builds next.