Q3 Investment Pipeline — Hugging Face, Perplexity, Poolside, and the Circularity We're Not Calling Circularity

Let's tally what we're actually proposing to spend before Friday's board call, because I don't think the deck captures it cleanly.

$12.9 billion for Hugging Face. Company revenue: roughly $150 million annualized. That's an 86x multiple on a platform that hosts 2.5 million models and 950,000 datasets and, by its own founders' account two years ago, turned down $500 million from us because they didn't want to compromise platform neutrality. We're now offering roughly 26 times that rejected number for the whole thing. Something changed. It wasn't the revenue.

Separately, we're in talks to put money into Perplexity's new round at a $30 billion-plus valuation, more than 50% above the $20 billion mark it set less than a year ago. Perplexity's annualized revenue has gone from under $250 million in January to $750 million now, which sounds impressive until you note the growth is partly a function of Perplexity Computer running workloads that consume the exact compute we sell. We are proposing to fund the customer that buys our chips so it can grow fast enough to justify a valuation that makes our investment look prescient in hindsight. This is not a new structure. It rhymes with the Poolside licensing arrangement we signed earlier this year for $6 billion, and it rhymes with arrangements Lucent and Nortel used to run with their fiber customers in 1999, right up until the customers stopped paying and the vendor got left holding both the receivable and the write-down.

Nobody on this committee needs the history lesson. I'm including it anyway because the committee keeps acting like the circularity concern is a media narrative rather than a balance sheet fact. It is a balance sheet fact. When we fund Perplexity's round and Perplexity spends part of that money on our GPUs, we have converted an equity check into revenue recognition. That is not fraud, it's disclosed, Reuters and The Information have both reported the mechanics — but it is also not organic demand, and every analyst modeling our forward multiple off "AI is reaching its inflection point" is modeling off a demand curve we are partially manufacturing ourselves.

Here's the part of the memo that should actually worry people: none of this is happening in a friendly rate environment. Fed Chair Warsh spent Friday's Jackson Hole keynote walking through a 3.7% core PCE print and getting September hike odds from 35% to 57% in the space of one speech. The 2-year yield touched its highest level since July. Gold got taken down 3% in a session, which tells you money was scrambling for dollars, not safety. We are proposing to write an $12.9 billion check and lead a $30 billion-plus round in a week when the risk-free rate just repriced higher and the market flinched. Our own cost of capital is not falling. Our targets' valuations are rising anyway. That gap has to be justified by something other than "AI has reached its inflection point," because that phrase, respectfully, is not a discount rate.

I want to flag the legal overhang on Perplexity specifically, since it's been treated as a footnote. Amazon sued Perplexity over Comet's agent accessing Amazon accounts and placing orders without going through Amazon's own channel. The Ninth Circuit vacated the initial injunction on narrow computer-access-law grounds, which is a win, but the case is back at the district court and Amazon can still pursue contractual claims. As a minority investor we inherit some exposure to that; had this been a full acquisition instead of an equity round, we'd own the litigation outright. Someone should confirm the round structure protects us from anything beyond our check size, because right now the deck treats this as resolved and it isn't.

Silver hit $70 an ounce intraday this week for the first time on record. I mention this not because it's our problem but because it's a useful gut check on where capital is actually fleeing right now — into hard, uncorrelated, can't-be-diluted assets, at the same moment we're asking this committee to write checks into equity rounds priced at 40x and 86x revenue for companies whose growth partly depends on continued access to our own hardware. The market outside this building is pricing debasement risk. Inside this building we're pricing eighteen more months of exponential AI demand as though it's underwritten by something other than our own spending.

None of this means don't do the deals. Hugging Face gets us distribution and a genuine moat around the open-source layer that a competitor could otherwise own. Perplexity gets us a seat at a company Bezos and SoftBank already backed, ahead of a 2028 IPO its CEO has said will happen regardless of how OpenAI or Anthropic list. Those are real strategic rationales and I'd defend them in front of the board.

What I won't defend is presenting the revenue these investments generate as evidence of external demand when a meaningful share of it is our own money moving through someone else's income statement and back into ours. If a hike lands in September and growth-stock multiples compress the way they did the last three times the 2-year moved 10 basis points in a week, I'd like this committee on record as having known exactly how much of our growth story was self-funded before that happened, not after.