The most interesting number in technology last week was not a benchmark score or a parameter count. It was about $125 billion, which is roughly the cash Alphabet was sitting on when it decided, on Monday, June 1, to go out and sell stock. Companies with $125 billion in the bank do not normally pass the hat. Alphabet passed the hat for $80 billion, and by Tuesday the demand was heavy enough that the hat had to be enlarged: the offering priced at $84.75 billion, which the financial press described as the largest equity raise in the history of a publicly traded technology company. When a business that prints money starts selling pieces of itself, that is not really a financing decision. It is a tell.

The tell got louder when you saw who was buying. Warren Buffett's Berkshire Hathaway, an institution that spent the last decade handing cash back to shareholders through buybacks, agreed to a $10 billion private placement: $5 billion of Class A stock at $351.81 a share, $5 billion of Class C at $348.20, per Alphabet's own filing. Berkshire has been quietly building an Alphabet position since the back half of 2025, but a $10 billion anchor check from the most famous technology skeptic in American finance is a different kind of statement. Either Buffett has decided the AI buildout is the safest bet on the board, or he has decided it is going to be expensive enough that he wants to own the toll road. With Buffett it is usually the toll road.

Read the structure of the deal and the picture sharpens. The $80 billion came in three parts: a $30 billion underwritten public offering, split between mandatory convertible preferred shares and common stock; a $40 billion at-the-market program that does not even begin until the third quarter; and Berkshire's $10 billion. Here is the part the press releases glided past, and that TechTimes caught: roughly $30 billion of that at-the-market tranche is earmarked not for graphics chips but for the tax bill Alphabet owes when employee stock grants vest. So the clean line that this is $84.75 billion "for AI infrastructure" is, at the margins, a little bit of housekeeping wearing a moonshot costume. Only the underwritten piece and the Berkshire money are explicitly compute.

Still, the reason Alphabet needs the money at all is the genuinely staggering part. Sundar Pichai told investors the company is compute-constrained in the near term, which is corporate for "we cannot buy chips fast enough." Alphabet's 2026 capital expenditure guidance now stands at $180 to $190 billion, roughly double the $91.4 billion it spent in 2025. This from a company that generated about $174 billion in operating cash flow over the trailing twelve months and grew revenue 22 percent year over year to $110 billion in the first quarter. Sit with that. A business throwing off $174 billion in cash, growing at 22 percent, has concluded that its own cash flow is not enough, and that selling equity while the stock is high beats running the balance sheet thin. The buildout has outgrown the richest cash machine in the index.

Alphabet is not alone, which is the whole point. The four hyperscalers — Microsoft, Amazon, Alphabet, and Meta — spent roughly $131 billion on capital expenditure in the first quarter of 2026 alone. Amazon led at about $44 billion, Microsoft around $31 billion in cash capex, Alphabet $36 billion, Meta $20 billion, and every one of those figures is roughly double what the same company spent a year earlier. Stitched together, the four are guiding toward something close to $700 billion of capex for the full year, a figure Epoch AI notes has quadrupled since GPT-4 shipped. The bull case is simple and not stupid: demand is real, the revenue is growing, and agents that run for hundreds of steps eat compute the way furnaces eat coal.

The bear case lives in the footnotes. Microsoft has acknowledged that roughly $25 billion of its 2026 capex is not more capacity at all but component price inflation — the same chips, costing more. And the firm has conceded that about two-thirds of its spending now goes to short-lived assets, meaning GPUs and CPUs that depreciate over a few years rather than the decades a building lasts. That is the uncomfortable arithmetic underneath the trillion-dollar talk: a meaningful slice of history's largest infrastructure boom is being poured into hardware that will be obsolete before the financing that paid for it matures. You do not have to believe the boom is fake to notice that the accounting is doing some heavy lifting.

If you want to know where the $700 billion is going, the answer spent last week on a stage in Taipei. At GTC Taipei on June 1, Jensen Huang used his Computex keynote to remind everyone that NVIDIA no longer merely sells the shovels — it is busy designing the gold. The headline was Nemotron 3 Ultra, announced Monday and released Thursday, June 4: a 550-billion-parameter mixture-of-experts model with 55 billion active parameters, a hybrid Mamba-Transformer design, a million-token context window, and an explicit purpose, which is orchestrating agents that plan, call tools, and grind through hundreds of turns without burning a fortune in tokens. NVIDIA claims up to five times the throughput and roughly 30 percent lower cost on agentic work.

The benchmark deserves the asterisk your readers expect. Artificial Analysis scored Nemotron 3 Ultra at about 48 on its Intelligence Index, which makes it the most capable open-weights model to come out of an American lab, ahead of Gemma 4 and OpenAI's gpt-oss. But Artificial Analysis ran that evaluation in partnership with NVIDIA ahead of release, so it is a vendor-blessed number, and even taken at face value it sits behind China's Kimi K2.6 at 53.9. The cleanest way to put it: NVIDIA just shipped the most intelligent open model in America, which is also the second-most-intelligent open model on Earth. The frontier of "open" is increasingly being set in Hangzhou, not Santa Clara.

The same keynote produced Cosmos 3, which NVIDIA calls the first fully open omni-model for physical AI. The pitch is that a single mixture-of-transformers system can reason about a scene, generate a world, and predict an action, collapsing what used to require three separate models into one, and Huang framed it with characteristic restraint as the "big bang of physical AI." Cosmos 3 landed on Hugging Face with a Nano variant small enough for a workstation GPU and a coalition of robotics and video labs — Runway, Black Forest Labs, Skild AI and others — attached to it. Strip the theater and the strategy is plain: every model NVIDIA gives away for free is a model that makes you want to rent more of its chips. The picks-and-shovels company has learned to seed the gold rush it profits from.

While the giants raised money and the chipmaker raised expectations, the data-infrastructure layer did the thing companies do when they are afraid of being swallowed: it consolidated. On June 1, Fivetran completed its all-stock merger with dbt Labs, first announced back on October 13, 2025. George Fraser stays on as chief executive, dbt's Tristan Handy becomes president, and the combined company claims around $600 million in revenue and more than 100,000 data teams. The interesting news was not the merger, which everyone knew was coming, but what shipped alongside it: dbt Core version 2.0, which open-sources the new Fusion engine under an Apache 2.0 license, and a preview tool called dbt State that the company says cuts pipeline compute by 30 percent or more by rebuilding only what changed.

The honest way to read that merger is as defense. Two companies last privately valued at $5.6 billion and $4.2 billion do not combine from a position of overwhelming strength; they combine because the hyperscalers are absorbing the data plumbing one feature at a time, and standing alone is getting dangerous. Fivetran helpfully published a "2026 agentic AI readiness index" the same week, reporting that 60 percent of enterprises are pouring millions into agentic AI while only 15 percent have the data foundation to run it safely. Note that as what it is — a vendor survey from a vendor that sells exactly the foundation in question — and the directional point still holds. The money is rushing into agents faster than the plumbing can carry it.

The plumbing is also where the alliances quietly rearranged. On June 1, OpenAI's GPT-5.5 and GPT-5.4, along with its Codex coding agent, reached general availability on Amazon Bedrock, a month after entering limited preview. Pricing matches OpenAI's first-party rates, and the practical consequence is strange to say out loud: Amazon's cloud is now the only place where you can run both OpenAI and Anthropic frontier models behind one API and compare them on the same console. OpenAI, Microsoft's closest partner, is now selling through Microsoft's largest cloud rival, on the back of an Amazon commitment that grew from $38 billion last November to $50 billion by February. Two days later, on June 3, OpenAI widened Codex further beyond pure coding. The frenemy map of this industry is being redrawn in real time, and the redrawing is worth more attention than any single model release.

One smaller item, for completeness and no more than a line: Vercel added xAI's Grok Imagine Video 1.5 to its AI Gateway on June 3, an image-to-video model with synchronized audio in a single pass, which Elon Musk showcased the next day with an AI-generated trailer. Useful for developers, not a market-mover.

So what actually happened last week? Strip away the keynotes and the model cards and you are left with a single, coherent story told in the language of money. The companies with the most cash decided it was not enough. They issued equity, leaned on debt, and pointed the proceeds at hardware that depreciates faster than the bonds funding it. The most disciplined investor alive wrote a $10 billion check to ride along. The chipmaker gave away the software that sells its chips. The independents merged for cover. And the rival labs started selling through each other's clouds. Every one of those moves is rational on its own. Taken together, they describe an industry spending against a future it is certain of and cannot yet prove, on a scale that has quietly passed the point where operating profits can pay the bill.

That is the tell. Not the benchmarks, which are vendor-marked and already contested by a Chinese model nobody in Washington wants to talk about. The financing. The financing is the part that has to be true, because you cannot fake a settlement date, and Alphabet's underwritten shares settle around June 4 with the depositary tranche following on the fifth. Snowflake Summit runs straight through this coming week, which means more agentic-data announcements are queued and more readiness indexes will be waved around. Watch them if you like. But the question that decides who is right is not which model tops a leaderboard next week. It is whether anyone, anywhere in this arms race, blinks on the spending. So far, with $700 billion on the table and Buffett at the felt, nobody has.