The Weekly Tycoon

Stocks rose and crypto didn't: the S&P gained 1.15% while ether fell 6.85%. The 10-year ended almost where it started after a scary Monday, so rates weren't the seller this week. Crypto's own funds were.

1. MACRO + GEOPOLITICSThe 10-year touched 5.33%, and buyers met it there

   The 10-year Treasury yield hit 5.33% on Monday, its highest since 2002, per Trading Economics. Then the Fed's minutes landed Wednesday and didn't help: most officials think another hike will likely be appropriate by year end. Fed governor Christopher Waller added that he expects more hikes if the data come in as expected, though not necessarily at back-to-back meetings. And yet the yield finished the week at 5.24%, a touch lower than it started. The reason is the buyers. The week's 10-year and 30-year auctions both drew strong demand at multi-decade highs, and bonds rallied after the 30-year sale. The plain read: money still costs more than it has in a generation, but this week it stopped getting more expensive, because people showed up to lock in those yields.

   Go deeper · Federal Reserve · FOMC minutes

2. BITCOINCrypto funds lost $1.25 billion in a week stocks went up

   Bitcoin ETFs gave back $681.1 million last week, ending a three-week run of inflows, per SoSoValue data reported by The Block. Ether funds did worse, with $542.1 million out, their worst week since January. Solana ETFs posted their first losing week after 14 straight weeks of inflows. Add it up and about $1.25 billion left the three, in a week when the S&P rose. Bitcoin itself held up better than its flows, ending near $82,500. That's just under $83,000, the level analysts quoted by Decrypt call the average ETF buyer's cost. So the typical fund holder sits roughly at break-even, and holders at break-even tend to sell into bounces. That's why the price barely moving matters more than the outflow: the funds sold and the market absorbed it.

   Go deeper · The Block

3. AIA trillion-parameter model you'll be able to download

   Mistral released Large 4 on Tuesday, a model with 1 trillion parameters that it plans to publish as open weights by the end of October. Open weights means you can download the model and run it on your own servers instead of renting it through someone's API. Until then, its API costs $1.36 per million input tokens, per VKTR. A day earlier, Reflection AI announced Beam, an open model built for coding and agents, under the Apache 2.0 license, which allows commercial use. Its weights are also due this month. Two frontier-scale open models in one week puts a lid on what closed labs can charge for everyday work. It lands just as Anthropic gets ready to host IPO investors at its headquarters on Oct 14.

   Go deeper · VKTR

4. CHIPS + ENERGYTSMC's sales flattened out at a very high level

   TSMC, the company that makes the chips for Nvidia, Apple and AMD, reported September revenue of NT$511.86 billion. That's up 54.6% from a year earlier. It's also down 0.6% from August, which is the part almost nobody quoted. One flat month isn't a verdict, since chip output moves in steps as new factory lines come on. But it's the last hard number before Thursday's third-quarter report. TSMC guided that quarter to $44.6 to $45.8 billion in revenue. The thing to read on Thursday is the spending plan for next year, because TSMC only builds factories for orders it believes in. A bigger plan says the AI chip build has another leg. A flat one says the plateau in September wasn't a fluke.

   Go deeper · TSMC

5. DIGITAL MEDIATikTok brought its ad network to American apps

   At Advertising Week on Monday, TikTok said it's expanding the TikTok Ad Network into the US. That's the system that places TikTok's advertisers inside other companies' apps, and it already reaches almost 400,000 third-party and developer apps worldwide, per Social Media Today. TikTok also launched Agentic Leads, an AI that collects sales leads through conversations in DMs and on advertisers' own websites. For anyone with a free app, a new buyer for your ad space is more competition for your inventory, which is how ad rates go up. For anyone selling, look at the full picture: TikTok now wants to run the ad, the conversation, the lead and the checkout. That's a whole sales funnel, rented from one platform.

   Go deeper · Social Media Today

AI's software got cheaper this week. Its power bill now runs 20 years.

Here's the deal. On Tuesday, Google signed for 3.59 gigawatts of electricity from Constellation Energy, per Bloomberg. About 890 megawatts of it comes from nuclear plants, on a 20-year contract. The rest is grid power on a 15-year deal. Constellation's stock jumped more than 13% on the news. That same Tuesday, Mistral put out a trillion-parameter model it plans to let anyone download for free. Put those two side by side and you get the shape of the AI business right now: the software gets cheaper by the month, and the physical stuff underneath it gets locked up by the decade.

Here's how the Google deal works. It isn't new reactors. Constellation will spend more than $4.3 billion upgrading 11 existing reactors in Illinois, Pennsylvania and New Jersey so they put out more power, with the first of it arriving in 2028. Upgrading plants that already have permits and grid connections skips the slowest parts of building new ones, and Google pays for that speed with a very long promise. The push came from the grid itself. PJM, the operator that runs power for much of the eastern US, has proposed that data centers bring their own power or risk being cut off at peak demand. Its capacity prices are up more than 11-fold since 2024. Power stopped being a utility bill. It's now a supply contract you fight for.

Now zoom out. A 20-year power contract is a bond in disguise: a fixed promise that runs for decades, and its value depends on what money costs. This was the week money cost the most in a generation, with the 10-year touching 5.33% on Monday. The Fed's minutes, out Wednesday, named one reason in plain text: "competition for capital from heavy private debt issuance to finance the development of artificial intelligence (AI) infrastructure." The AI build is now big enough to show up in the central bank's explanation of what borrowing costs everyone else.

Meanwhile the software layer is moving the opposite way. Mistral's Large 4 costs $1.36 per million input tokens through its API, and Reflection's new Beam model ships under a license that lets anyone use it commercially. Anthropic, which has invited institutional investors to its headquarters on Oct 14 ahead of an expected IPO, will be pitching a business whose product now has free, downloadable rivals near the frontier. That doesn't make the labs bad businesses. It makes their edge a moving target, while a nuclear contract signed this week keeps paying into the 2040s.

Follow the money and it keeps landing on the long-lived stuff. Constellation's jump was the market paying up for contracted power. TSMC's September sales were up 54.6% from a year ago. And JPMorgan's analysts say bitcoin miners have been net sellers of about $1.8 billion of coins, cashing out to build AI data centers, per The Block. The miners are a small version of the whole story. They're trading a volatile asset priced every second for long, steady leases with tech companies. Even the people who mine bitcoin are picking the 20-year contract.

That's also the frame for crypto's rough week. Bitcoin, ether and Solana funds lost about $1.25 billion while the S&P rose. When cash pays over 5% and the hottest trade in the market is long-dated infrastructure, the money left in fast trades gets pickier. Ether took the hit, down 6.85% on the week. Bitcoin gave up much less and held near the price where the average ETF buyer got in. The selling came through the funds, and the price mostly absorbed it.

So here's the idea to carry into the week. AI now runs on two clocks. The scarce things, power, chip factories and sites, get signed up for decades and priced off the bond market. The abundant thing, intelligence itself, gets cheaper every few weeks and is increasingly free to download. Those are two different businesses, and most of the confusion about the AI trade comes from treating them as one.

Why it mattersThe AI build now runs on 20-year contracts priced off a 10-year yield near a 24-year high, so the cost of money sets the price of AI's physical layer while the models keep getting cheaper.

Sources · BNN Bloomberg · Google and Constellation · Federal Reserve · FOMC minutes · VKTR · Mistral Large 4 · Stocktwits · Anthropic investor day · The Block · JPMorgan on crypto flows · TSMC · September revenue · Trading Economics · US 10-year

1. Dealers got stuck with 2.5% of the 10-year auction

   Wednesday's 10-year Treasury auction sold $39 billion of notes at a 5.300% yield. The detail is in who bought. Primary dealers, the big banks that have to soak up whatever nobody else wants, were left with just $984 million. That's about 2.5% of the competitive sale. Indirect bidders, the bucket that includes foreign central banks and big funds, took about 80%. A market that's running out of buyers leaves the dealers holding the bag. This one didn't, which helps explain why the 10-year ended the week a little lower than it started. The tell to watch: if dealers have to step in heavily at a future auction, the line of buyers has moved.

   Go deeper · TreasuryDirect · auction results

2. Apple named AI agents as the reason to lock down the Mac

   On October 2, Apple told developers it's tightening Full Disk Access on macOS, the permission that lets an app read your files, mail, messages and browsing history. Its reason was unusually direct: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially." Granting it will soon take "very explicit user action." It came days after a report that Meta's Muse app read private messages without explicit permission. For anyone building desktop agents or AI tools that read local files, onboarding just got a step harder. The bigger signal: platform owners have started writing the rules for agents, and they're writing them for safety first and convenience second.

   Go deeper · TechCrunch

3. YouTube quietly opened ad money to political explainers

   YouTube's ad guidelines page has a short October 2026 entry most creators won't read. Certain content of public interest, with "congressional proceedings" and "debates over ongoing government actions" as the examples, will now be eligible to earn ad revenue. In September, the same page opened ads to graphic gaming content in a video's first eight to fifteen seconds. Put together, YouTube is loosening the brand-safety filter that used to strip ads from anything heated. The timing matters, with the US midterms on November 3. Commentary, civics and explainer channels, a big faceless category, just got a better shot at full ads in the busiest news season of the cycle.

   Go deeper · YouTube Help · ad guideline updates

Q. What's an open-weight model, and when does running my own beat paying for an API?

An open-weight model is one whose trained brain, the big file of numbers called weights, is published for anyone to download. You can run it on your own machine or a rented server, tweak it, and build on it. A closed model, like the top ones from OpenAI or Anthropic, only works through the company's API: you send text, they send back an answer, and they bill you by the token.

Running your own tends to win in three cases. When you have steady, heavy volume, because a server rented by the hour can cost less than millions of API calls. When the data can't leave your hands, like client files or anything medical. And when you need the model to stay exactly the same, since API models get updated or retired on someone else's schedule.

The API wins almost everywhere else. A model like Mistral's trillion-parameter Large 4 needs serious hardware, and someone has to keep it running. For most solo builders, the practical setup is both: build on an API, keep your prompts and test cases portable, and treat open weights as your exit door and your price check. When a free model can do the job, it caps what any vendor can charge you.

Sources · VKTR · Mistral Large 4 · Let's Data Science · Reflection Beam