The flood that came down the Bhotekoshi on Wednesday morning did not come from rain. Nepal's Department of Hydrology and Meteorology logged under seven millimetres in Rasuwa in the preceding twenty-four hours. The United States Geological Survey first recorded the event as an earthquake, then corrected itself: what its instruments caught was a glacial collapse and debris flow, and the shaking was the flood, not the cause of it.

The working reconstruction, based on Planet Labs imagery reviewed by Nepal's National Disaster Risk Reduction and Management Authority, is that a mass of ice and rock came off a slope roughly twenty kilometres northeast of the Rasuwagadhi border crossing, dammed the Lhende river, impounded water behind itself, and then released. Rijan Bhakta Kayastha, a Kathmandu University climate researcher who has spoken with scientists on both sides of the border, says it will take several days to establish whether a glacial lake burst as well. Authorities in Nepal and China have not yet agreed on a cause.

They do agree on the arithmetic. Nepal's police have recovered 177 bodies and deployed 3,318 officers. The disaster authority puts 826 people missing. Fourteen staff of the Rasuwa Customs Office are unaccounted for, out of thirty-nine. Around sixty houses went into the river near Trishuli Bazaar. Two motorable bridges are gone, which is why casualties are being flown to Chakradev Hospital in Bidur rather than driven to Trishuli.

And then there is the grid. The Ministry of Energy, Water Resources and Irrigation's preliminary inventory reads like a list of everything Nepal did right for ten years. The 111 MW Rasuwagadhi plant. The 22 MW Chilime. The 78 MW Sanjen Khola. The 20 MW Langtang Khola, the 5 MW Mailung Khola, the 6.42 MW Upper Mailung A. Downstream in Nuwakot: the 24 MW Trishuli, the 14.1 MW Devighat, the 60 MW Upper Trishuli-3A, the 37 MW Upper Trishuli 3B, the 15.6 MW Middle Trishuli Ganga, the 25 MW Sotar. Under construction and hit anyway: the 216 MW Upper Trishuli-1 and the 120 MW Rasuwa Bhotekoshi. A 25 MW solar plant lost three blocks and a substation.

Add it up and it is about 754 megawatts across roughly a dozen projects on two rivers. Not all of it was generating — Upper Trishuli-1 and Rasuwa Bhotekoshi are still being built — and that distinction matters enormously to the grid and not at all to the lenders. Nepal's total installed capacity at the close of fiscal 2025/26 was 4,120 megawatts. One morning touched something close to eighteen percent of it.

The most consequential damage may be the least photogenic. The Trishuli 3B hub substation was hit, which has disconnected projects from the national transmission system that were not physically damaged at all. This is the standard grim lesson of grid failure: the turbines are not the fragile part. The single point where fourteen of them come together is.

The timing is cruel in a specific way. Nepal spent the last decade converting itself from a country with scheduled winter blackouts into a net electricity exporter. In fiscal 2025/26 it sold 3.877 billion units to India and Bangladesh for Rs 29.32 billion, up sixty-eight percent, at an average of Rs 7.56 a unit, while cutting imports by nearly a third. That export business runs on monsoon surplus. This is the monsoon.

What makes Wednesday a story about the next flood rather than only this one is the sensor that is not there. On 8 July 2025, a glacial lake outburst came down this same Lhende river. Afterwards, Nepal and China agreed to share cross-border information on floods, landslides and glacial lake risk. On 1 August 2025, officials in Tibet's Nyalam County told counterparts in Nepal's Sindhupalchok District that they would pass along warnings by WeChat and phone. The head of Nepal's disaster authority, Dinesh Prasad Bhatta, said at the time that the country needed "an institutional and written mechanism for information sharing, not one based on personal contacts."

Thirteen months later there is still no signed agreement. The issue was raised again during a Nepali delegation's visit to China in May. On Wednesday, with the water already in Nuwakot, an official told the Kathmandu Post that Nepal was "seeking real-time information" from Chinese authorities and from research centres, because "there is no one at the Nepal-China border point."

A 2024 study in *Science of the Total Environment* assessed 545 glacial lakes in the Sino-Nepal corridors and rated 144 of them highly susceptible to outburst. India, China, the United States and the World Bank have all offered support for relief and reconstruction. Reconstruction is the expensive part. The warning was the cheap part, and it is the part nobody bought.

Georgia's Public Service Commission approved Georgia Power's contract with OpenAI on Wednesday. The deal, filed in July for a project in Effingham County, covers 3,200 megawatts of new demand. OpenAI pays the full cost of the infrastructure built to serve it and has committed up to 1,000 megawatts of flexible demand response, meaning the utility can throttle deliveries to the site during system peaks rather than build generation for them.

Read that structure carefully, because it is now the template. The interesting move in American power regulation over the past eighteen months has not been about clean energy or capacity markets. It has been the invention of a class of customer so large that it gets its own body of law.

Wisconsin's commission ruled in April that any customer with more than 100 megawatts of peak demand must subscribe to enough generation to cover it, and must pay for the generators built on its behalf even if it walks away early. Michigan set the same 100-megawatt threshold and ordered Consumers Energy to file a case for each large-load customer demonstrating that no residential or commercial ratepayer is subsidising it. Indiana regulators worked through a contract that scales Amazon to 2,400 megawatts by 2032, with purpose-built generation and depreciation schedules attached. American Electric Power says it has signed agreements covering 69 gigawatts of incremental load through 2030.

Every one of those orders is, functionally, an insurance policy written against the same event: that the load does not show up. Minimum bills, exit fees, subscription requirements, no-subsidy certifications. Regulators have become genuinely sophisticated at pricing the risk that a data centre gets cancelled.

What none of them has priced is the other case.

Georgia Power's announcement leads with the affordability number. Its portfolio of large-load contracts is projected to deliver roughly $950 million a year in savings beginning in 2029, worth $2.847 billion across 2029 to 2031, which the company translates as $180 a year for a typical household using 1,000 kilowatt-hours a month — up from the $102 it promised in December. Indiana Michigan Power is running the same play, proposing $59 million in reductions and a three-year rate freeze on the strength of large-load revenue.

This is a defensible argument. Fixed grid costs spread over more kilowatt-hours do lower everyone's unit cost, and a customer paying its own interconnection is a better customer than one that isn't. But note the tense. The commitments are stated in dollars per year beginning in 2029. The construction, the interconnection queue, the transformers and the turbine orders happen between now and then, and they happen at 2026 prices in a market where every utility in the country is bidding for the same equipment.

The household is being asked to hold a three-year receivable against a counterparty whose revenue model is younger than the transformer's delivery lead time. That may well pay off. It is worth being honest that it is a trade, and that the party on the other side of it did not sign anything.

The numbers first, because they are extraordinary and because they are not the story. Nvidia's second quarter of fiscal 2027, ended 26 July, produced $96.22 billion of revenue against a $92.07 billion consensus and a $91.0 billion guide midpoint — up 106 percent year on year and 18 percent sequentially. Non-GAAP earnings were $2.22 a share. Gross margin held at 75.0 percent. Data centre revenue was $89.0 billion, or about 92 percent of the company. The third-quarter guide is $108 billion against $103.8 billion expected, which puts Nvidia through the $100 billion quarterly mark on a single product line.

Then the chief financial officer, Colette Kress, told the call to expect roughly 70 percent revenue growth in fiscal 2028, excluding any Chinese data centre revenue. The street had 45 percent. Nvidia's own framing of the constraint is not demand but supply: the top five cloud providers will spend close to $800 billion this year, perhaps $1.3 trillion next, and the company says it sees no deceleration in their plans.

Here is the disclosure worth clipping. Nvidia has formed strategic partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500 billion of third-party capital for AI infrastructure. Separately, it is working with SB Energy at the PORTS-Pike Technology Campus in Ohio to secure land, grid interconnection and building shells in advance of the compute that will occupy them.

A chip company does not need a consortium of private-credit sponsors and an option on a brownfield site in Pike County. A project-finance sponsor does. What Nvidia disclosed on Wednesday is that the binding constraint on its growth has moved off its own wafer allocation and onto its customers' ability to raise money and pour concrete, and that it has decided to solve both.

The obvious comparison is telecom vendor financing in 1999, and it is worth making the comparison precisely rather than glibly, because the differences are real. Lucent and Nortel lent their own balance sheets to buyers who could not otherwise pay, booked the sale, and ate the default. Nvidia is syndicating other people's capital to buyers — Microsoft, Alphabet, Amazon, Meta — who are among the most cash-generative enterprises in history. Those four spent $166.0 billion on capital expenditure in the June quarter alone, up 87 percent year on year and 27 percent from March. Combined hyperscaler capex has risen 272 percent over ten quarters. This is not a subprime borrower.

But syndication moves risk; it does not delete it. UBS puts hyperscaler capex at roughly $4.1 trillion across 2026 to 2028, more than three times the previous six years combined, and estimates Amazon, Alphabet and Microsoft will spend about 102 percent of their cloud revenue on capex this year. Bank of America sees $860 billion in 2026 and a path to $1.2 trillion in 2027, and used the word "unfathomable" in its own note.

When the supplier has to organise the financing, secure the land, and pre-book the electrical interconnect, the supplier has concluded that the buyers cannot do it fast enough on their own. That is a statement about the pace of the buildout, and it is also a statement about where the marginal dollar of risk has gone: out of Santa Clara, into a fund, and eventually into a substation in Ohio that has to be energised on schedule.

President Trump declared a national emergency on Wednesday over foreign cyber and security threats to the American bulk-power system, prohibiting the purchase or installation of certain foreign-made electric equipment and software judged to pose sabotage or unauthorised-access risk. The Energy Secretary is directed to impose conditions on equipment already in service and to publish operating rules under the emergency.

The legal instrument is familiar; a version of it appeared in the first term, and the same statutory family has been used this year for steel, copper, semiconductors and pharmaceuticals. The substance is worth separating from the framing.

First, an emergency declaration about hardware is really a declaration about replacement time. Governments do not invoke extraordinary powers over a component they can re-source in a quarter. Large power transformers are custom-built, weigh hundreds of tonnes, and have been quoted at multi-year lead times through the current buildout. An order that says *stop installing this* is only meaningful if the answer to *install what instead* is difficult, and the difficulty is the point.

Second, the order explicitly excludes facilities used for local distribution of electric energy. The bulk-power system is the transmission backbone; distribution is the last few miles, the part with a pole outside your house. The reasoning is defensible — the backbone is where a single compromised device can propagate — but it means the perimeter has been drawn around the equipment the federal government can most easily reach, not necessarily around the equipment closest to the customer.

Third, this administration has now issued two grid orders in five months pointing in opposite directions on the same question. April's executive order used emergency authority to keep retiring plants running, which the White House says preserved 17,000 megawatts. Wednesday's uses emergency authority to restrict what can be installed. Both are answers to the same underlying condition: American demand for electricity is rising faster than American capacity to build the physical objects that carry it.

That condition is not a foreign plot. It is an order book.

The document is a Request for Information, solicitation HSIvoter26, posted Tuesday by Immigration and Customs Enforcement through Homeland Security's Investigations and Operations Support office in Dallas. Responses are due Friday, 28 August, at 6:00 p.m. UTC, submitted through Microsoft Forms. Performance would be in Arlington, Virginia. An RFI is market research; it obligates the government to nothing.

What it describes is a supplier who would "identify, request, and acquire publicly available voter registration and history files from U.S. jurisdictions," delivering them to Homeland Security Investigations to support "fraud detection and data segmentation activities." The draft statement of work asks the contractor to obtain at least one public voter registration file each year from every jurisdiction the government names, plus separate voter history files for federal general, primary, runoff and special elections, with the capacity to scale to all fifty states, the District of Columbia, and the territories.

Much of the reaction has focused on the scope. The word to sit with is *annual*.

There is a meaningful difference between a demand and a subscription. A demand for a state's voter file is an event: a state can refuse it, a secretary of state can litigate it, a court can enjoin it, and the news cycle can hold it. Standing infrastructure that ingests every jurisdiction's registration roll once a year and every federal election's history file as it is published is not an event. It is a capability with a maintenance schedule, and it stops being news after the second delivery.

The other phrase worth noting is "data segmentation," which is a marketing term. Segmentation means dividing a population into groups by attribute in order to treat the groups differently. In a fraud investigation you are looking for specific records. In segmentation you are sorting everybody.

HSI is the criminal investigative arm of ICE, and its portfolio genuinely includes fraud. The context is nonetheless the context: the administration has spent the year asserting that the voting system has been compromised, with the President claiming in July to have evidence of Chinese efforts to obtain voter data from as many as eighteen states. The Supreme Court on Monday allowed enforcement of an executive order building lists of citizens eligible to vote and restricting mail ballot eligibility. The midterms are in November.

The procurement asks that acquisition follow "neutral and auditable procedures." That is a real safeguard and worth holding the eventual contractor to. It is also worth noticing that the audit trail proposed here covers how the data was obtained. It says nothing about what the segmentation is for.

A reader who works in state election administration and has fielded one of these requests: we would like to hear from you.

Overnight into Thursday, Israeli artillery shelled a hilltop near Beit Jinn in the western Damascus countryside, and soldiers raided and searched a house in the village of Taranja in the Quneitra countryside, according to witnesses and Syrian state broadcaster Alikhbariah. In Lebanon, the National News Agency reported air attacks on al-Mansouri in the Tyre district and Sarbine in Bint Jbeil, strikes on Beit Lif and Yater, artillery on Wadi Zibqin, and large explosions at Bayt al-Sayyad, Baraachit and Majdal Zoun. Demolition of property continued, including at Hula in the Marjayoun district.

All of this occurred inside a diplomatic process that is, by its own terms, succeeding. Syria's foreign minister, Asaad al-Shaibani, met Israeli intelligence officials days ago in talks brokered by the United States. Israel and the Lebanese government have announced an American-mediated deal under which Israeli forces leave southern Lebanon in exchange for Hezbollah's disarmament — a step one US official called "huge." Hezbollah has refused direct talks and its leader has said the group will not disarm.

The pattern here is now old enough to name. A framework agreement in this theatre does not end fighting. It establishes a ceiling and a tempo, and both parties operate confidently just beneath it, because the framework's existence is itself valuable to each of them and neither wants to be the one who voided it.

The demolitions are the detail to watch, more than the airstrikes. Bombing is a statement about the next week. Systematically demolishing structures across a belt of villages is a statement about the next several years: you clear ground you expect to be looking at for a long time. The Lebanese debate over deploying the army south, and over what disarmament could even mean operationally, is happening against buildings that are being removed while it happens.

The wider clock reads six months. Trump launched the war with Iran in late February forecasting four to five weeks. A June memorandum of understanding meant to open a path out has lapsed. American stockpiles of key munitions have drawn down and the administration has shifted from strikes to economic pressure. Hezbollah's March escalation was framed as a response to the killing of Ali Khamenei on the war's first day, and Iran has insisted that the fighting in Lebanon must be settled as part of any deal — which places the least tractable file at the centre of the negotiation.

Meanwhile the meter runs elsewhere. Reuters calculates that countries affected by these conflicts produced about 45 million barrels a day based on 2025 output, more than 43 percent of global supply, with Saudi Arabia rerouting crude to the Red Sea and Gulf exporters moving cargoes quietly out of a Strait of Hormuz that remains shut. A war that was supposed to take five weeks has become a permanent feature of the price of everything.

Online reservations for pass holders opened this morning for the 83rd Venice International Film Festival, which runs on the Lido from 2 to 12 September under Alberto Barbera. The opening film, premiering in competition in the Sala Grande on the second, is *Ink* — directed by Danny Boyle, written by James Graham, starring Jack O'Connell, Guy Pearce and Claire Foy.

Graham's stage play of that title, which won him an Olivier and ran at the Almeida and then in the West End and on Broadway, was about Rupert Murdoch's 1969 purchase of a failing broadsheet called *The Sun* and the year that followed, in which Murdoch and his editor Larry Lamb rebuilt it around a simple and unforgivable question: what do readers actually want, as opposed to what we have decided they should want?

It is a very good play, and part of why it is good is that it refuses to let you feel superior. Lamb's newsroom is not stupid. It is doing market research the establishment press considered beneath it, and it is right about the findings, and the findings are corrosive. The play's second act is the bill arriving.

So: the most prestigious festival in Europe, an institution that has spent two decades turning itself into the launchpad for American awards season while insisting on its curatorial soul, is opening with the origin story of the popular press. Barbera has a sense of humour about his own building, which is one of the more likeable things about him.

The rest of the lineup confirms the mood. The lifetime achievement Golden Lions go to George Clooney and Ellen Burstyn — one a movie star who became an institution, the other an actor who has spent fifty-five years refusing to. Masterclasses at the Match Point Arena from 3 to 9 September bring Chloé Zhao with her cinematographer Łukasz Żal, plus Luc Besson and Luca Guadagnino, which is a lineup that covers roughly the entire available spectrum between austerity and appetite. Greta Scarano and Nicolas Maupas host both ceremonies.

There is a version of the Venice preview that lists the competition titles and handicaps the Golden Lion. We will run that in a week. For now the interesting thing is the choice of overture. Film festivals open with a statement about themselves, and this year's statement is a film about people who discovered that if you ask an audience what it wants and then give it to them, you will win, and you may not like what you have won.

Every publication is running that experiment right now, including this one. It seems worth watching somebody dramatise the first time it was run properly.

Consider what happened in Ohio this week, because it is genuinely impressive and almost nobody described it that way.

Nvidia disclosed that it is working with a developer to secure land, grid interconnection and empty building shells at a campus in Pike County *before* the computers that will occupy them exist. That requires forecasting electrical demand years out, contracting for it, and accepting the carrying cost of an empty steel box in the meantime. On the same Wednesday, Georgia's regulators approved a utility selling 3,200 megawatts to a company whose flagship product did not exist eight years ago, with the customer paying the full cost of the infrastructure and pre-committing a third of its load as interruptible.

These are acts of long-range planning of a very high quality. They happen for one reason. If the electricity is not there on the promised date, an identifiable balance sheet loses an enormous amount of money, and everyone involved knows exactly whose.

Now look at the Lhende.

An automatic gauging station on a glacial lake, a water-level sensor, a satellite uplink and a siren downstream is not a frontier engineering problem. The technology is forty years old. Nepal already runs monitoring on Tsho Rolpa and Imja. Instrumenting the high-risk lakes across the Sino-Nepal corridors and running them for a decade would plausibly cost less than the 111-megawatt Rasuwagadhi plant lost on Wednesday morning.

So why doesn't it exist? The usual answer is poverty, and the usual answer is lazy. Nepal exported Rs 29.32 billion of electricity last year, and the developers on the Trishuli corridor are commercial entities with real capital at stake.

It does not exist because the sensor has to sit twenty kilometres inside another country, and there is no instrument by which a Nepali hydropower developer can pay a Chinese county government to install one and be contractually entitled to the reading. The two states agreed in principle after the last outburst on this same river on 8 July 2025. Local officials in Nyalam County offered to relay warnings by WeChat. It was raised again in Beijing in May. Nothing has been signed. Thirteen months, one river, two floods, no gauge.

Call it what it is: a missing market. The good being traded is *advance notice*, the buyers are identifiable and solvent, the sellers are identifiable, and no contract connects them. Everything else in this business has been solved by exactly this move. Catastrophe reinsurance turned hurricane forecasting into a product. Parametric drought insurance pays on a rainfall index. Georgia Power turned uncertainty about a data centre's arrival into an exit fee. When someone can be paid to reduce a risk, the risk gets measured.

The 2024 assessment in *Science of the Total Environment* found 144 of 545 lakes in the Sino-Nepal corridors highly susceptible to outburst. On the Bhotekoshi and Trishuli alone there are now thirteen damaged projects with named owners, named lenders and named insurers. That is not a charity case. That is a customer list.

So: not another pledge of reconstruction support, though the reconstruction money will be needed and should come. The lenders on the Trishuli corridor should jointly fund an upstream monitoring network as a condition of refinancing, and somebody — ICIMOD, a reinsurer, a development bank willing to be commercially imaginative — should build the legal vehicle that lets them pay across the border for it.

The alternative is the arrangement we have now, in which the most advanced planning apparatus on earth can conjure a substation in rural Ohio for a machine that has not been designed, and the same civilisation cannot get a float switch onto a lake in Tibet that has already flooded once. That is not a failure of capability. It is a failure of ownership, and ownership is the one thing this era knows how to invent.

The Federal Reserve chair gives his debut Jackson Hole speech tomorrow, and Goldman's chief US economist has already told clients not to expect policy guidance. That is not a criticism of Goldman's read. It is an accurate description of Kevin Warsh's stated project, which is to restore the Federal Reserve's credibility by having it say considerably less.

The case for that project is stronger than its critics allow, and I want to put it properly before disagreeing with it. Forward guidance in 2021 was not merely wrong; it was a trap. Having told the world that rates would stay at zero into 2024, the committee found that reversing itself cost more than the guidance had ever bought, and it reversed too slowly as a direct result. A commitment device becomes a hostage. Warsh's view — that a central bank which explains its framework and then simply responds to data is more credible than one issuing quarterly promises it may have to break — is a defensible, even conservative, position with a long lineage.

But it has a bill, and the bill has arrived.

The Fed's preferred inflation gauge came in Wednesday at 3.7 percent for July, unchanged from June and up from 2.9 percent when the war with Iran began in late February. Core rose 0.2 percent on the month. Petrol is around $4.10 a gallon. Inflation has now exceeded the 2 percent target for more than five consecutive years. Markets moved to price a 40.1 percent chance of a hike of at least 25 basis points in September, up from about 36 percent before the print; the Kalshi ladder reads roughly 32 percent hike, 67 percent hold. Boston's Susan Collins said Tuesday the Fed will need to raise soon absent clear progress. The committee is split, and everyone can see it is split.

Meanwhile the thirty-year Treasury touched a nineteen-year high earlier this month, and the Treasury Secretary responded by doubling the size of long-bond buybacks — an intervention unusual enough that the dollar wobbled on debasement talk before recovering. Wall Street analysts have been fairly direct that a lack of clarity from the chair is itself part of what is holding long rates up.

That is the mechanism people keep missing. Forward guidance was never a gift to traders. It was the central bank absorbing forecasting risk that would otherwise sit in the term premium. Take the absorption away and the risk does not vanish; it gets repriced and paid for by whoever holds duration, which in practice means mortgage borrowers, corporate issuers and the Treasury itself. The buybacks are the invoice. A government that must intervene in its own long bond because its central bank has stopped explaining itself has not eliminated a cost. It has moved the cost to a different department of the same government.

So here is what I would like from the podium tomorrow, and it is not a September signal. Warsh should not forecast. He should specify. He should say what inflation path, over what horizon, would compel this committee to raise, and what path would not. That is not guidance about an outcome; it is a description of a reaction function, which is precisely the thing his own framework says a credible central bank owes the public. Silence about intentions is a legitimate policy. Silence about criteria is just opacity wearing the costume of discipline.

Five years above target is not a communications problem. At some point it is a preference, revealed. If it is not a preference, the chair has one speech tomorrow to say so in terms specific enough to be wrong.