Upside is a weekly podcast designed to look behind the headlines that will affect European venture, startups and investing.

Below are the notes from this week’s episode. Episode links above to tune in and stream wherever you pod.

Quick Hits Before the Main Event

A run through the week’s other headlines:

Broadcom missed its own estimates by $200 million on a $35 billion quarter and fell 3%, even after AI chip revenue tripled. The real tell, per Mads, was the two-year guide: Broadcom expects that AI revenue to double next year and double again the year after.

Matt Clifford, who wrote much of Britain’s AI strategy, is joining Anthropic as Managing Director of International Affairs. Pri’s take: “the most effective place to shape AI policy is inside a US lab’s international affairs function.”

Apple’s new CEO John Ternus has his first big event on 9 September, with a foldable iPhone and a Siri rebuilt on Google’s Gemini expected. Pri called renting the model layer, for about a billion dollars a year against a $4.6 trillion market cap, “the most rational decision in the industry.” Andrew called the foldable “an also ran.” I disagree: Apple always lets others innovate first, then ships its own polished version.

Nvidia confirmed it’s buying Hugging Face for $12.9 billion. Mads thinks it hands Nvidia real control over open source distribution, since Hugging Face is the shelf developers browse before picking a model.

Wayve cars, safety driver still up front, are now running on Uber in London, the same week Uber cut 3,300 jobs. Pri’s read: “their bet is on the licensing model, they’re building the software, not the car.”

And Andrew welcomed the UK’s new £100 million sovereign AI competition, paying British companies to solve public-sector problems instead of handing out grants: “the government needs to write contracts to solve problems that ultimately will help the taxpayer.”

Meta Puts a Price on Your Data

The headline story this week is Meta’s Muse Spark 1.3, and it’s a clever piece of business model design. Mads laid it out: pay the private rate, roughly one to four dollars per million tokens, and Meta never trains on what you send it. Or switch off the data retention policy, let Meta train on what you send, and pay up to 95% less.

Mads’ framing stuck with me: “what we have seen over the years is that free, or nearly free, is a class of its own. Incredible businesses have been built on the back of monetising data.” His comparison was Robinhood, free trading paid for by selling order flow. Free AI inference, paid for with your data, is the same trade in a different outfit.

I’m sceptical. I don’t trust Meta with my data and said so on the pod. But I’m not the customer. The pitch lands hardest with small businesses bolting AI into a product without eating a $500 a month bill, and Andrew’s addition: Meta’s tooling will likely be good enough that businesses skip the consulting layer open source usually required.

There’s a second thread worth pulling on the models themselves. OpenAI’s new Astra and Anthropic’s Fable 5.1 both landed this week, and Mads flagged something in Astra that matters more than the benchmark scores. Today’s reasoning models write their thinking out in words before answering. Astra reportedly skips that for parts of its reasoning, computing in a numeric form instead, more efficient because the model isn’t constantly translating its own internal state into English and back. The catch: reading a model’s chain of thought is one of the cheapest safety tools we have, and it stops working once the thinking is never expressed in words. Mads isn’t usually in the alarmist camp on AI safety, but on this one, “I have a little bit of sympathy.” His ask: labs should keep a way to reconstruct the reasoning after the fact.

Why Bond Yields Are Suddenly Interesting

I’ve become a bit obsessed with bond markets. Treasuries, guilts, whatever your country calls them, this is the interest rate at which governments borrow, and it sets the floor under mortgage rates, business borrowing and consumer credit. Ten-year yields across the G7 are up three to four percentage points in five years. Germany went from being paid to borrow to its highest yield since 2011. Japan hit a 30-year high. The UK now pays the most in the group, over 5%.

Scale matters too. The global bond market, roughly $160 trillion, is bigger than global equities at $155 trillion. This week gave a genuinely odd signal: tankers got hit, oil spiked, and instead of the usual flight into bonds for safety, investors sold them.

Mads doesn’t think this is a crisis yet, but there’s something real underneath it. Two record borrowers are drawing on the same pool of savings at once: governments running deficits with no real attempt at control, the US alone spends about a trillion dollars a year just servicing its debt, more than defence, while the AI build-out needs capital on a historic scale too. His read on where it ends: either someone gets serious about spending, or AI delivers growth big enough to outrun the debt, or something eventually breaks. Major Western economies had debt below 50% of GDP not long ago. Now it’s around 100%, closer to 120% in the US, and Japan is over 200% and still climbing.

Andrew’s addition was about duration: if AI-driven borrowing persists, the era of very low, long-term rates may simply be over for a while, not from a shock but because demand for capital isn’t going anywhere.

The Circular Financing Chart Got Bigger

About a year ago a Bloomberg chart showing AI’s circular financing went viral, the arrows connecting OpenAI, Nvidia and Oracle back to each other. Vendor financing has existed forever, and some of the original panic was overcooked. But the chart has changed shape. By my own count, it’s now 29 entities and 61 capital flow lines, with Nvidia in the middle of roughly $750 billion of deals connected to its own customers.

The real change since last year is structural. Commitments are turning into guarantees, and guarantees are quietly moving off balance sheet. Meta borrowed $27 billion for a single Louisiana data centre, but only a fraction of that shows up on its own balance sheet, in bonds rated A-plus and maturing 2049, long after that generation of chips retires. Oracle sold $18 billion of bonds in a single day and is still fighting a bondholder lawsuit filed back in January. SB Energy and Nscale both filed for IPOs this week on backlogs rather than revenue: SB Energy has $439 billion of backlog against $139 million of first-half revenue and a $3.2 billion loss, and Nscale’s $103 billion headline is a 5.7-year contract backlog, not turnover.

Mads’ read captures where the risk has moved. Venture capital at the earliest layer is a small slice of this. Above it sits real debt, underwritten by insurers, pension funds and private credit betting that billions in value will still be there years from now. Nvidia is now offering to guarantee some of that residual value itself, which Mads thinks is smart given how well GPU values have held up, but it’s still a bet on history repeating.

Predictions

Oracle is rated BBB- by S&P, one notch above junk, and the market prices it more generously than that implies. My prediction: if OpenAI has any kind of stumble, Oracle gets downgraded to junk by year end. Oracle has tied itself tightly to the OpenAI ecosystem, and as Mads put it, “if Anthropic takes OpenAI to the cleaners and OpenAI ends up not performing, Oracle will be in trouble.”

Deals of the Week

Mads: iPronics, $125 million Series B. A Valencia based silicon photonics company spun out of the local university in 2019, backed by Nvidia, Bosch and the European Innovation Council fund. Its rack mounted optical switch lets GPUs talk over light instead of fixed copper, cutting idle time. A proper European deep tech play, one layer down from the frontier model race Europe has largely given up on.

Pri: Instinct, $350 million at a $2.5 billion valuation. A Silicon Valley AI assistant living inside iMessage that books restaurants, flights and tickets, monitoring for cancellations and rebooking while you sleep. Free for consumers, backed by Benchmark and Index, and it wants your email and card connected to do it. As Pri put it: “we are the product.”

Notable Quotes

Mads Jensen, on Meta’s data-for-discount pricing:

“What we have seen over the years is that free, or nearly free, is a class of its own. Incredible businesses have been built on the back of monetising data.”

Priyanka Savjani, on Nscale’s $103 billion “contracted revenue” figure:

“This metric represents future revenue tied to multi-year lease agreements for data centre capacity, much of which has yet to be built.”

Andrew Sherlock, on the UK’s £100 million sovereign AI competition:

“There are too many grants out there and it motivates the wrong behaviours within startups. The government needs to write contracts to solve problems that ultimately will help the taxpayer.”

Mads Jensen, on Oracle’s exposure to the OpenAI ecosystem:

“If Anthropic takes OpenAI to the cleaners and OpenAI ends up not performing, Oracle will be in trouble.”