UBS just published its global real estate bubble index. Across 23 cities, real house prices rose an average of 0.5% over the past year.
Half a percent. After inflation, that's roughly nothing.
But "average" is the most misleading word in that sentence. Pull it apart and these 23 cities don't look like one market at all.
Seoul is the standout, up more than 10% in real terms. The reason isn't subtle: the AI-chip boom pushed a wall of money into Korea.
Canada is the biggest retreat. Vancouver and Toronto fell 10% in real terms, against a backdrop of Canada tightening rules on foreign buyers.
Most US cities also declined in real terms, with one exception: San Francisco — the AI capital.
Germany looks flat to gloomy, with Frankfurt and Munich down nearly 4% after inflation.
See the pattern? This cycle isn't splitting by country. It's splitting by whether a city caught the AI money. Seoul and San Francisco aren't rising because their housing is better. They're rising because their industries happen to be standing where the money is landing.
Now the bubble list, which is more interesting than the rankings.
Of 23 cities, UBS flags only two as "high" bubble risk: Zurich and Tokyo.
Zurich is a rate story — ultra-low Swiss rates have kept capital parked in property for years, and prices reflect it.
Tokyo is something else. Over the past seven years, real prices there have risen roughly 50%, driven by rising buyer leverage and a nascent pickup in wages.
Read that again: 50% over seven years, not one. Anyone looking at "Tokyo is rising" and feeling the urge to move should check the time scale first.
Then there's London, which I think is the most useful sample for ordinary buyers.
Prices fell 1% year on year. Sounds mild — until you account for the fact that London is down about 15% in real terms since 2021. The report's phrase is "stretched affordability," which translates to: it's not that nobody wants to buy, it's that they can't.
The UK as a whole looks the same. Lloyds puts the average home at £298,441 in September, almost exactly where it was a year earlier — while the average five-year fixed mortgage has hit 6%.
Prices stopped. Payments didn't. That's the real texture of a high-rate housing market.
And there's a structural shift underneath all of this that matters more than any single city.
McKinsey Global Institute calculates that global household wealth grew by $40 trillion in 2025. About 57% of those gains came from stock markets. Property contributed just 15%.
Here's the sharper line: in the US, this July, equities overtook real estate as the largest component of household wealth — the first time since World War II.
That tells you something about the function of housing. As a wealth-creation vehicle, at least in the aggregate statistics, it has handed the crown to equities.
But I wouldn't jump to "housing is finished." UBS itself notes that underlying urban demand remains solid, that rents have broadly kept pace with income over the past five years, and that London rents are still near record highs because of a persistent supply shortage.
So my read is this: housing is shifting from an asset that makes you rich to an asset that keeps inflation from eating you. Those are completely different holding logics. The first is about leverage and appreciation. The second is about cash flow and rent.
So, do you still want to buy property for investment purposes?
Source: UBS Global Real Estate Bubble Index 2026 (via MoneyWeek, 2026-10-16) | Lloyds UK house price data | McKinsey Global Institute
Moonshot just shipped Kimi K3.1 Pro: 3 trillion parameters, 1 million tokens of context.
Two numbers, two different claims. Parameter count is capacity — how much it can think with. Context window is memory — how much it can hold at once.
The capability list is more telling than the spec sheet. Four areas were called out: coding, 3D creation, financial research, and long-horizon agent work.
Put those together and you get one sentence: this model is aimed at jobs that take a long time to finish.
Long-horizon agents are the hard case. Not "answer this question," but "take this task, break it into steps, run the tools, check the output, notice when it's wrong, and fix it" — across dozens of turns. That's exactly where context windows earn their keep. If the model forgets what it did three steps ago, the whole run collapses.
Which is why I care more about the second model in the release: K3.1 Pro Highspeed, described as the same intelligence at roughly two to three times the output speed.
Speed is the underrated variable. A large part of agent cost is waiting — waiting for it to reason, to call a tool, to retry. Doubling or tripling throughput doesn't just improve the experience; it cuts the cost per completed task.
My own experience with these tools is simple: capability decides what it can do, speed decides whether you'll actually use it. A model that writes good code but takes five minutes will get used once a week. The same model at one minute gets used every day.
There's a second thread here worth flagging. Reports describe this as a 3T-class open-weight release. If the weights really are open, the story isn't "a new model" — it's "3 trillion parameters is now deployable on your own infrastructure." For small teams building vertical products, that changes the cost structure, not just the benchmark.
Now the cold water, because this category has a habit of overpromising.
First, parameters and context are capacity metrics, not quality metrics. A million-token window means nothing if the model loses the thread at 800,000. The only honest test is needle-in-a-haystack at scale, plus real multi-step tasks.
Second, "3D creation" is the easiest claim to inflate. Being able to drive Godot and Blender is not the same as producing 3D assets someone will pay for. There's an art pipeline and an engineering pipeline between those two things.
Third, no pricing was announced — and pricing is what actually decides whether teams switch. The last two years of price wars have reset everyone's mental anchor. A stronger model at the wrong price is a demo.
So my read: the weight of this release isn't "most powerful." It's the combination — long-horizon tasks, faster output, and open weights showing up in the same launch. That's the combination that moves agents from demo to daily tool.
Whether it replaces what you're using now is not a conference-stage question. Take your three most common tasks, run them once, and look at the bill and the clock.
@ajassy Cybersecurity has become one of the primary concerns for organizations and enterprises adopting AI, and it may well remain the most critical factor in the future.
Microsoft just walked away from a data centre project in Essex. Not because of money. Because the grid couldn't promise power before 2030.
That project had been publicly framed as a step toward making the UK an AI leader. It died on an interconnection queue.
This is the part of the AI story that gets less attention than chips and models: in Europe, the binding constraint is increasingly electrons and paperwork, not capital.
The numbers are not subtle. The IEA puts grid connection queues in major European data centre hubs — Frankfurt, Paris, Dublin — at seven to ten years. Seven to ten. A model generation lasts eighteen months. You can't build a business on infrastructure that arrives four model cycles late.
Then there's price. The Draghi report in 2024 found EU industrial electricity costs running two to three times US levels, and gas four to five times. In April, OpenAI paused its UK "Stargate" project in Tyneside, citing high UK power prices. When electricity is the input, the cost curve of intelligence is set by the cost curve of power.
And then there's politics, which is where this gets really hard.
Ireland once had data centres consuming roughly a fifth of national electricity — more than all its urban households combined. New data centres there are now expected to bring their own power. Denmark has an emergency plan that puts households first and large data centres last in the supply order.
The result: in the first four months of 2026, more than seventy data centre projects in Europe were rejected or restricted — more than in all of 2025.
I find this genuinely interesting because it inverts the usual narrative. Everyone assumes Europe's AI gap is about compute, talent, or risk appetite. Europe has talent. It has research. It has the AI Act, an "AI Continent" action plan, and a stated ambition to be the third pole. What it doesn't have is a grid that can say yes on a useful timeline.
So the constraint has moved from "can we build the model" to "can we get the interconnection agreement." That's a different kind of bottleneck — slower, more political, and much harder to solve with money alone.
For anyone building or investing around this, three implications stand out.
First, siting decisions are now power decisions. The question isn't "which city has good fibre," it's "where can I get firm capacity before 2030."
Second, the value migrates to whoever can deliver power: utilities, grid operators, behind-the-meter generation, long-duration storage, and the boring engineering firms that upgrade substations.
Third, Europe's constraint creates an export opportunity elsewhere. If European projects keep slipping, the compute goes to the US, the Gulf, and Southeast Asia — and the companies that learn to build fast in those markets will have a structural advantage.
The uncomfortable conclusion: Europe may end up regulating AI faster than it can power it.
That's not a technology problem. It's an infrastructure problem wearing a technology costume.
Four US crowdfunding portals got suspended in a single month. The register doesn't tell you why.
Here's what's actually on the record: FINRA's funding portal register shows four Rule 9552 suspensions within about a month — BRITE Investments and Community Bond from September 21, Dentist Fund from September 25, and MustrdSeed Portal from October 5.
Rule 9552 is the procedure for failing to provide information, or failing to keep it current. That's it. The register doesn't publish the underlying facts of each case, so I can't tell you whether these were sloppy filings, missed deadlines, or something worse. Nobody can, from the public record.
And that's the part that bothers me.
Crowdfunding's entire pitch is access: small companies raise from ordinary people, ordinary people get to own a piece of something early. The pitch works because the platform sits in the middle and does the vetting. When the middle layer goes dark, the investor finds out last.
Think about how you'd actually discover this. You invested in a Reg CF deal eighteen months ago. The company is fine. The platform isn't. Did anyone email you? Probably not. Did the portal's app keep working? Maybe. Would you know to check FINRA's register? Almost certainly not — and why would you? You didn't pick the portal because of its compliance filing history.
That asymmetry is the whole problem. Issuers are supposed to check their portal counterparty before routing an offering. Investors are supposed to check... what, exactly? The deal page, which the portal wrote.
I've been broadly optimistic about regulated crowdfunding. The structure is sound: a registered intermediary, mandated disclosure, capped raise sizes. But structures fail at the edges — and the edge here is a portal that stops filing while its deals stay live.
Three things I'd want if I were investing in this market. First, a status field on every deal page: is the portal currently in good standing, yes or no, updated automatically. Second, a notification obligation — if a portal gets suspended, every investor with a live position gets told, in plain language. Third, a visible audit trail of the last compliance filing, the way a fund shows its NAV date.
None of that is exotic. It's the minimum for a market that asks ordinary people to trust an intermediary they'll never meet.
Regulation Crowdfunding turns ten this month — the SEC's advisory committee marks it on October 27. Ten years in, the rules around who can raise and who can invest are getting more sophisticated. The plumbing around what happens when a portal fails is still catching up.
So here's my takeaway, and it's unglamorous: before you wire money into a Reg CF deal, spend ninety seconds on FINRA's funding portal register. Not the pitch deck. The register.
The four suspensions aren't a scandal. They're a reminder that in crowdfunding, the platform is the product — and platforms can quietly go away.