Syz the moment

A great FT article on how OpenAI partners amassed $100bn debt pile to fund its ambitions.

The Commitments: $250B in cloud compute from Microsoft $38B from Amazon 36 gigawatts of contracted capacity All tied to a total deal value up to $1.8 trillion HSBC’s estimate: OpenAI will owe ~$620B per year in data-center rent once everything ramps… and only a third of that capacity is online by 2030. 🔢 The Math (and the Problem) By 2030: Cumulative rental costs: $792B (→ $1.4T by 2033) Projected free cash flow: $282B Cash from Nvidia/AMD: $26B Undrawn debt: $24B Liquidity: $17.5B Even after stacking every possible dollar, there’s still a $207B hole — plus the $10B safety buffer HSBC thinks they need. 💥 And here's where it gets tricky 👇 HSBC’s model already assumes everything goes right: 3B OpenAI users by 2030 (44% of all adults outside China) Paid conversion rising from 5% → 10% 2% share of global digital ads $386B in annual enterprise AI revenue Even under that fantasy scenario, OpenAI still can’t pay the bills. HSBC’s suggested “solution”? OpenAI may need to walk away from its data-center commitments and hope Microsoft/Amazon “show flexibility.” Translation: The economics don’t work — unless everyone politely pretends the contracts aren’t real. And yet this is the company anchoring a $500B Stargate project and driving hundreds of billions in AI infrastructure spending. If this is what the best case looks like… imagine the base case. My take: be very careful with AI plays which are asset-heavy. They might disappoint in terms of shareholders' returns in the years to come. Do you remember the Telecom bubble? The long-term winners have been the asset-light companies. The asset heavy companies never recover. Source: Hedgie on X, FT

👉 But none of it shows up on Meta’s balance sheet. How? Meta shifted the entire project into a joint venture: 🔹 Meta owns 20% 🔹 Blue Owl Capital owns 80% 🔹 A holding company (Beignet Investor) issued $27.3B in bonds, mostly bought by Pimco 🔹 Meta will rent the data center starting in 2029 And here’s the kicker: the lease is structured to qualify as an operating lease, not a finance lease — letting Meta avoid listing the giant asset and the massive debt. But peel back the layers and things get messy: 🔥 Meta runs the data center 🔥 Meta carries the risk of cost overruns 🔥 Meta guarantees the full value of the bonds if they don’t renew 🔥 Yet Meta insists it doesn’t “control” the venture enough to count it on the books Even the Wall Street Journal called it “artificial accounting.” 🧩 It’s part of a bigger trend: Tech giants want unlimited AI infrastructure… 🚫 …but they don’t want the debt that comes with it. Morgan Stanley estimates the industry could need $800B in off-balance-sheet financing by 2028. Meta may not be borrowing on paper — but economically, this is debt with extra steps. What do you think: smart financial engineering or a red flag in disguise? Source: Hedgie

The "Google bets" basket (Alphabet, Broadcom and Celestica) just hit a new ATH while the "Open AI" basket (Nvidia, Softbank & Microsoft) has been hut hard since the end of October. Source: Steno Research, Macrobond, Bloomberg

It’s meaningfully better at everyday tasks like working with slides and spreadsheets. The new AI tops coding benchmark, leading in key tests like SWE-bench Verified at 80.9%, Terminal-bench 2.0 at 59.3%, and OSWorld at 66.3%, beating models from Google and OpenAI in coding, agent tasks, and computer use. It features a 200K token context window, uses far fewer tokens for the same work, and costs much less at $5 per million input tokens. Developers can now access it through APIs, apps, and platforms like Amazon Bedrock and GitHub Copilot, with engineers noting its strength on complex bugs. Source: CNBC-TV18

Source: Bourbon Capital @BourbonCap

Source: Koyfin @KoyfinCharts

Google's AI chips - TPUs, or tensor processing units - are having a moment. These semiconductors were used to train its latest genAI model, Gemini 3, which has received rave reviews, and are cheaper to use than Nvidia's offerings. 🚀 But here's the real reason Google invented the TPU Back in 2013, Google ran a simple forecast that scared everyone: If every Android user used voice search for just 3 minutes a day, Google would need to double its global data centers. Not because of videos. Not because of storage. But because AI was too expensive to run on normal chips. So Google made a bold move: 👉 Build its own AI chip - the TPU. 15 months later, it was already powering Google Maps, Photos, and Translate… long before the public even knew it existed. ⚡ Why TPUs Matter GPUs are great, but they were built for video games, not AI. TPUs were built only for AI. No extra baggage. No wasted energy. Just raw efficiency and speed. That focus paid off: TPUs deliver better performance per dollar Use less energy Are faster for many AI tasks And with each generation, Google doubles performance Even Nvidia’s CEO, Jensen Huang, openly respects Google’s TPU program. 🤔 Then why don’t more companies use TPUs? Simple: Most engineers grew up with Nvidia + CUDA, and TPUs only run on Google Cloud. Switching ecosystems is hard — even if the tech is better. ☁️ The Bigger Picture: Google’s Cloud Advantage AI is crushing cloud margins because everyone depends on Nvidia. Google isn’t. It owns the chip and the software stack. That means: ✔️ lower costs ✔️ better margins ✔️ faster innovation ✔️ and a defensible advantage competitors can’t easily copy Some experts now say TPUs are as good as or even better than Nvidia’s best chips. 🔥 The Punchline Google didn’t build TPUs to sell chips. It built them to survive its own AI growth. Today, TPUs might be Google Cloud’s biggest competitive weapon for the next decade. And the moment Google fully opens them to the world? The AI infrastructure game changes. Source: zerohedge, uncoveralpha