THE AI TRADE - WHO BUILDS, WHO BORROWS, WHO BENEFITS

Chipmakers at the bottom - they take the profits. The builders in the middle who set it all up and take the risks. The users at the top - who receive the benefits, mostly on free subscriptions. 

This guide maps the six layers, follows the money and sets out the risks, from interest bills to China and the power grid.

Logic flow

Fast Track

The "AI trade" is a chain of businesses stacked on top of each other, from chipmakers at the bottom to customers at the top, and we can think of each layer as steps in a supply chain. The chipmakers are the first to profit: they sell a scarce product and are paid in the sales / delivery step. The builders in the setup step take the risks: buildings, fast-ageing chips and a growing pile of debt. The software tool-makers and end-users at the top get the benefits: powerful tools, often on free subscriptions, that are at the same time disrupting some of the world's most profitable businesses. 

For now, far more is being spent on setup than customers are paying, with borrowed money to fill the gap. From where a company sits in the chain, you can tell where its troubles are likely to start.

Introduction

For most of us, AI means a chatbot on a phone. For financial markets, it means a national building programme: the four largest American technology companies plan to spend roughly $730bn on chips, servers and data centres in 2026, up from about $410bn the year before.

Those same stack of companies make up about a third of America's main stock index, so "the AI trade" is shorthand for this whole market. But look deeper into the stack to see the companies earning money in different ways, carrying different kinds of risk and may not finally be the ones to gain the most. This article analyses the trade layer by layer, following the logic of the supply chain, and weighs the rewards and risks.

WHY IT MATTERS

Anyone with a pension or an index fund already owns a big slice of the AI trade through loans and stocks. But the build-out is spilling over into the economy more generally - from liquidity and interest rates, through to higher electricity bills. 

It promises to lift productivity for decades to come, but if investment overshoots the losses will spread, just as they did after the fibre-optic boom of the late 1990s.

CONTENTS COVER THE FOLLOWING

1. What markets mean by "the AI trade"
2. The stack: six layers from silicon to customer
3. Follow the money: where one dollar goes
4. Who gets what: profits, risks and benefits
5. The risk register
6. How stress travels, and what history teaches

1. What Markets Mean by "the AI Trade"

When ChatGPT appeared at the end of 2022, it looked like a clever piece of search software. It turned out to be the visible tip of a vast physical industry. Every question typed into a chatbot is answered by thousands of specialised chips in a warehouse-sized building that draws as much electricity as a small town. Each new, larger model needs more powerful chips, more sophisticated "AI Campuses" to house them, and more power of course.

That is why AI has become a story about capital spending rather than simply better code. Investors treat the companies involved as one big trade because their fortunes are all tied together along one supply chain. When cloud companies raise their spending plans, chipmakers rise. When doubts emerge about whether AI will earn enough long-term to justify the capex, hedge funds who've shorted the trade will take profits - almost everything on the chain falls at once.

IN PLAIN TERMS: WHAT IS AN "AI CAMPUS"?

An AI campus is not one building but a cluster of buildings: several spaces the size of warehouses are packed with chips, plus their own electricity power substation, then a cooling plant and of course fibre links to connect everyone on the same backbone, plus sometimes there will be a power station on site or next door. The biggest are planned at one to five gigawatts, such as the Stargate site at Abilene in Texas, built for OpenAI, or the planned Stargate UAE campus in Abu Dhabi.

IN PLAIN TERMS: WHAT DOES ONE GIGAWATT MEAN?

A gigawatt is a billion watts, about the output of a large nuclear reactor running flat out. Over a year, that is enough electricity on average for roughly two to three million homes in Europe. One gigawatt of AI campus holds several hundred thousand of the most advanced chips and costs somewhere between $35bn and $60bn to build and equip.

2. The Stack: Six Layers from Chip to Customer

The easiest way to understand the trade is to picture it as a building of six floors. Each floor buys from the one below and sells to the one above.

Layer	Examples	How it earns money
1. Chips	Nvidia, AMD, TSMC, SK Hynix	Sell processors and memory
2. Data centres and power	Builders, cooling firms, utilities	Equipment, space and electricity
3. Hyperscalers	Amazon, Microsoft, Alphabet, Meta	Renting computing power, plus their existing businesses
4. Neoclouds	CoreWeave, Nebius, Lambda	Renting out AI chips, mostly on borrowed money
5. AI labs	OpenAI, Anthropic, Mistral, xAI	Subscriptions, usage fees, advertising
6. Customers	Businesses and consumers	They pay for everything in the layers below in theory, many use the stack for free

At the bottom sit the chipmakers, led by Nvidia. In the quarter to July 2026 it sold $96bn of products and kept about 75 cents of every dollar after the cost of making them, an extraordinary margin for a manufacturer. Above them are the data centres: the buildings, cooling systems and grid connections that support th3e chips.

The hyperscalers are the giant cloud computing companies. What sets them apart is that they already have hugely profitable businesses, such as search advertising (Google), social media (eg Meta) and online shopping (Amazon), whose profits until recently paid for the AI build-out. The neoclouds are newer specialists that do only one thing, they rent out AI chips. They have no older business to fall back on and typically borrow heavily, often pledging the chips themselves as collateral.

The AI labs build the models. Their revenues are growing remarkably fast: OpenAI's annualised revenue was about $50bn in September 2026 and Anthropic's about $65bn in July. But they have also signed contracts to buy computing power worth hundreds of billions of dollars over the coming years.

Europe has its own contender. Mistral AI, founded in Paris in 2023. Mistral is Europe's leading model builder and the flagship of its ambition for "sovereign" AI that does not depend on American or Chinese suppliers. Its revenue run rate was estimated at about $400m in early 2026, with a target of more than $1bn for the year. ASML, the Dutch maker of the machines that make the chips, became Mistral's largest shareholder in 2025, and Mistral is building its own data centres near Paris and in Sweden for customers such as Airbus, BMW and the French military. It is small compared with the American labs, but it gives Europe a seat at the table and is an alternative it can control.

At the top are the customers, the only layer that brings genuinely new money into the system. Everyone else is paid out of what customers spend, or out of what investors and lenders advance in the hope that customers will generate enough returns.

3. Follow the Money: Where One Dollar Goes

Imagine a business paying an AI lab for access to its model. The lab passes much of that money down to the cloud company whose data centres run the model. The cloud company, in turn, has spent heavily on chips, buildings, electricity and interest, and must recover those costs over the life of the equipment. At the bottom, the chipmaker was paid in full when it delivered the chips.

How one dollar of customer money pays the stack

Using published cost estimates, one customer dollar looks like this. The lab keeps about 45 cents for its staff, research and sales, and passes about 55 cents to the cloud company. The cloud company keeps perhaps 10 cents as margin and to pay the loan. About 18 cents goes on buildings, networking and electricity. The remaining 27 cents goes on chips, and roughly 14 cents of that is Nvidia's gross profit. Nvidia's share is larger than everything the cloud layer keeps, and Nvidia is paid on delivery while the cloud company waits years for ROI return on investment.

Everyone in the chain is paid out of the customer's dollar, except the chipmaker, who is paid out of the builder's.

This neat split assumes the system is in balance. But it is not, not yet. The big four hyperscalers plan to spend about $6 on building this year for every $1 that customers currently pay the leading labs, and borrowed money and investors' capital fill the gap. Some readers will hear an echo of a Ponzi scheme. It is not a Ponzi scheme - a Ponzi pays early investors with later investors' money and has nothing real behind it, whereas this money buys real chips and buildings that earn real revenue. The closer comparison would be to a government deficit, spending today on the promise of growth tomorrow. The crucial difference is that no central bank stands behind the AI trade. A government can, in the last resort, have its debts bought with new money created by its central bank; a company must persuade lenders to keep lending... This is why the price of credit (interest rates) matters so much to this story.

4. Who Gets What: Profits, Risks and Benefits

Put the pieces together and a clear pattern appears with each part of the stack ending up with something different.

Profits at the bottom. Nvidia carries little debt, it sells a scarce product and is paid on delivery and as it carries little debt it keeps the margin. Whatever happens to AI later, the chips have already been paid for and NVidia keeps its considerable profits.

Risks in the middle. The cloud companies and neoclouds build and maintain the buildings, the ageing chips and, increasingly, the debt. Their lenders now take the chips and customer contracts as collateral. The trouble is that chips lose value quickly as more performant products come to market, and the worry is that contracts end before the loans are repaid. The labs, too, carry long purchase commitments that must be met from revenues that are growing fast, but are still only a fraction of what they promised to spend.

Benefits at the top. Users, meanwhile, are getting extraordinary tools for next to nothing. More and more people now ask an AI for an answer rather than searching the web, and AI agents can compare prices across rival shops in seconds. This nis all good for customers, but painful for incumbents. Investors fretted in October 2026 that growth in Google's search business, the engine of its profits, was slowing, and many commentators argue that the most profitable companies in the world are, in effect, disrupting themselves.

Economists call this consumer surplus: value that reaches consumers without showing up as profit for producers. We've been here before. Railways and airlines transformed economies while making many of their investors poorer. Fibre optic the same. The AI build-out may well create enormous value, but the question for investors is "how much of this new value will investors capture, and how much will simply flow to the people at the top of the stack?".

5. The Risk Register

The risks to the AI trade are easier to judge when they are set out side by side. Six stand out.

5.1 The interest bill

Debt has to be serviced whether or not the chips are busy. CoreWeave, the largest neocloud, earned $2.6bn of revenue in the second quarter of 2026 and paid $640m in net interest, roughly 25 cents of every dollar it took in, leaving it with a net loss of $626m*. In October some of its bonds yielded more than 15%, a sign that lenders fear they may not be repaid in full. The hyperscalers' interest bills are still small beside their profits, but they are growing: these companies have issued well over $200bn of bonds this year, and Meta reported negative free cash flow in the third quarter, its first since 2012.

5.2 Competition for capital

The build-out is moving savings out of financial markets and into physical things: concrete, copper wire, silicon.... It does so at a time of much competition for credit. Governments are borrowing heavily, and many companies must now refinance the cheap debt they took on during the Covid years of 2020 and 2021 but at much higher rates - this is the so-called "debt wall". When governments, refinancing companies and AI builders all compete for the same pool of savings, the price of money tends to rise. On top of this is inflation which central banks would traditionally tackle with higher rates. Several bond managers argue that this competition is one reason long-term interest rates have stayed high, which in turn raises the cost of the build-out itself.

If the AI trade fails, would governments step in, as they did for the banks in 2008? The banks were rescued because their collapse would have frozen payments across the whole economy; a failed data-centre landlord would not be the same. But Washington already treats AI as a strategic industry - for example, in 2025 it took a stake in the chipmaker Intel. Lobby pressure to support the largest players in a downturn cannot be ruled out. This is a question investors are bound to increasingly ask.

5.3 Chips that are overtaken by technology

Nvidia releases a new generation roughly every year. The cloud companies spread the cost of their chips over about six years. Analysts, notably the investor Michael Burry, argue that the useful life of a frontier chip is closer to two or three. If the critics are right, the true cost of AI is higher, and its profits lower, than company accounts show. (See the Deep Dive.)

5.4 China

American investors tend to overlook the competitor that may matter most. Barred from buying Nvidia's best chips, China has built its own: DeepSeek says it trained its latest model, V4, on Huawei's Ascend processors. Chip for chip, Huawei may still trail Nvidia, but China makes up much of the difference with scale and cheap power. Its models are close behind the leaders, by DeepSeek's own estimate three to six months, and a lot cheaper. In April 2026 DeepSeek priced its top model at about $3.50 per million tokens of output (roughly 750,000 words), against $25 to $30 for leading American models. If AI good-enough  for most purposes becomes cheap, the prices American labs can charge, and so the revenue meant to pay for the whole stack, will come under pressure.

5.5 Power, grids and geography

Much of the risk is passed on to the power system. American utilities plan to spend at least $1.4 trillion by 2030, with data centres the leading driver of new spending, and in the eastern United States data-centre demand has already pushed up household bills and provoked local moratoria. Europe is affected too. In 2023 data centres used nearly 80% of Dublin's electricity and a third or more in Amsterdam, London and Frankfurt... waiting times for a grid connection in these hubs now run seven to ten years. The EU wants to triple its data-centre capacity - this will require a rapid and major upgrade of the grids.

Geography adds a further danger. The Gulf states hoped to become AI hubs, with campuses such as the planned five-gigawatt Stargate for UAE. The war with Iran has shown the risk: cheap drones struck Amazon data centres in the UAE and Bahrain in March 2026, when Iran declared the Strait of Hormuz closed. Undersea cables run through the Gulf and the Red Sea, which carry most data traffic between Europe and Asia. With repair ships unable to reach them safely, a damaged cable could stay broken for months. on end When three Red Sea cables were cut in 2024, a quarter of the traffic between Asia, Europe and the Middle East was disrupted.

5.6 Will customers pay?

The whole stack rests on customers paying. ChatGPT had more than a billion weekly users by August 2026, but only about 50 million paid consumer subscriptions, roughly 5%. Is that a warning or an opportunity? It is both. The warning is that most people take AI like Google search for granted, something that should be free, and cheap Chinese models make charging harder still. The opportunity is that a billion regular users is an immense base. Advertising inside ChatGPT had reached about $1bn of annualised revenue by August 2026, and business customers already provide more than half of OpenAI's income. Google built one of the world's most profitable companies on a free product. Whether the AI labs can do the same will decide whether the top of the stack can pay for the floors beneath it.

6. How Stress Travels, and What History Teaches

The stack also predicts where trouble shows first. The weakest balance sheets, the indebted neoclouds and the labs that depend on fresh investment loans, are hit first because they need a constant resupply of money. The hyperscalers come next, through falling share prices and dearer borrowing, but at least they have profitable businesses to fall back on. The chipmakers come last, only when their customers finally decide to cut orders. Hence the market phrase that an AI sell-off moves "from the outside in".

History offers a warning and some comfort. In the late 1990s, telecom companies laid huge networks of fibre-optic cable for an internet they expected to grow explosively. This happened - the internet did transform the world, but the timing and the financing were off. Estimates vary, but perhaps 85% to 95% of that fibre was still unused after the bubble burst, and more than 60 telecom companies failed between 2001 and 2003. And yet the fibre survived and, a decade later, was carrying the world's video and smartphone traffic, just as its builders had imagined. A technology can be right while the investment in it is wrong, and the users, not the investors, can end up the winners.

Today's builders are far stronger than the telecom upstarts of 2000 of course, and today's customers are real too. Here are five signals worth watching:

• Neocloud bond yields, the earliest warning of stress at the stack's outside edge
• Hyperscaler spending plans and cash flow, which show whether the build-out is speeding up or slowing, and how it is being paid for
• AI lab revenue against compute commitments, including how many users conversion rates into paying customers
• The price of AI, both the rent for chips and the price of models - cheap Chinese competition is showing here
• Electricity prices and grid connections, the physical limit on how fast the build-out can proceed.

Bottom Line

The AI trade splits its rewards and risks unevenly. The chipmakers get the profits because they sell the startup product and can demand payment on delivery. It is the builders in the middle who take the risks as they are the ones carrying costly buildings, fast-ageing chips and growing debt piles. But the end-users and add-on software developers at the top get the immediate benefits by gaining powerful tools for little or nothing, while former incumbents are pushed out by members of this new industry.

Seen this way, market news reads differently. A slump in neocloud bonds indicates doubt - perhaps a cheap Chinese model threatening the prices that were expected to pay for the whole stack. Financial success or failure of the build-out will depend on people in sectors well beyond the technology sector - from the bond markets that determine mortgage, car loan and credit rates, to ordinary people paying their electricity bills. So the question to consider is that if most of the value flows to software startups and end-users, who is going to cover these enormous capital setup costs?

DEEP DIVE: TWO DEBATES BEHIND THE NUMBERS

How long does an AI chip last? Take a chip costing $40,000. Spread over six years, it adds about $6,667 a year to costs; over three years, about $13,333, and reported profit falls by the difference. Scale that up to the hyperscalers' estimated $425bn of server purchases in 2026 and the yearly cost is about $71bn at six years but about $142bn at three. Michael Burry estimates that overly long assumed lives could overstate the industry's profits by about $176bn between 2026 and 2028. The defence is the "value cascade" - a chip too slow for training new models can still earn money answering everyday queries. Rental prices for older chips will show who is right.

Is the money going round in circles? Some deals send money round in a loop. In September 2025 Nvidia said it would invest up to $100bn in OpenAI, OpenAI planned to fill new data centres with Nvidia chips, although Nvidia's chief executive later described this as an invitation rather than a firm commitment. Nvidia has also agreed to rent back any unused capacity from CoreWeave, one of its biggest customers. Critics call this "circular financing" and recall the telecom equipment makers of the late 1990s, who lent to customers who later failed. Defenders say it is simply how scarce capacity is secured, as when airlines pre-order aircraft or car hire companies block buy advance fleets. The test is whether, over time, the loop is closed by outside customers' money or only by the next round of equity investment and borrowing.

Related Articles

• Inflation, the dollar, gold and gold miners: how interest rates shape the value of long-term assets.
• France's debt and the bond market: why bond investors' appetite matters for governments and companies alike.
• Financial and Physical

Glossary

AI campus - A cluster of large data-centre buildings on one site, with its own power and cooling, built to train and run AI models.
AI trade - The group of investments linked to building and using artificial intelligence, which markets tend to treat as a single bet.
Annualised revenue - Recent monthly or quarterly revenue scaled up to a yearly figure; a snapshot, not a full year's results.
Capital expenditure (capex) - Spending on long-lasting assets such as buildings, servers and chips.
Circular financing - When a supplier invests in or lends to a customer, which then uses the money to buy the supplier's products.
Collateral - An asset pledged to a lender, which the lender can seize if the loan is not repaid.
Consumer surplus - The value users gain from a product over and above what they pay for it.
Debt wall - A cluster of loans and bonds falling due at around the same time, which must be repaid or refinanced.
Depreciation - Spreading the cost of equipment over the years it is expected to be useful.
Free cash flow - The cash a company has left after paying its running costs and its investment spending.
Gigawatt - One billion watts, about the output of a large nuclear reactor; now used to measure AI campuses.
GPU (graphics processing unit) - The type of chip used to train and run most AI models.
Gross margin - The share of revenue left after the direct cost of making the product.
Hyperscaler - One of the giant cloud companies, such as Amazon, Microsoft and Alphabet.
Neocloud - A specialist company that rents out AI chips, usually financed heavily with debt.
Sovereign AI - AI models and infrastructure that a country or region controls itself rather than relying on foreign suppliers.
Token - The small chunk of text, roughly three-quarters of a word, by which AI models measure and price their output.
Yield - The annual return a bond pays at its current price; a rising yield means lenders see more risk.

References

• Nvidia, results for the second quarter of fiscal 2027, 26 August 2026.
• BuilderMuse, hyperscaler capex guidance, October 2026; CNBC, Alphabet holds 2026 capex unchanged, 28 October 2026; CNBC, Amazon holds 2026 capex unchanged, 29 October 2026.
• Epoch AI, cost breakdown of a one-gigawatt AI data centre, May 2026; Investing.com, Bernstein on the cost of a gigawatt.
• Yahoo Finance/Bloomberg, OpenAI's annualised revenue, 8 October 2026; Sacra, Anthropic, OpenAI and Mistral revenue and funding, 2026.
• TechCrunch, Mistral raises $830m in debt for a data centre near Paris, 30 March 2026.
• CoreWeave, second-quarter 2026 results, 11 August 2026; Fortune, CoreWeave and neocloud debt, 16 October 2026.
• CNBC, AI debt and hyperscaler free cash flow, 27 October 2026; TwentyFour Asset Management, hyperscaler issuance and the Treasury curve, 3 September 2026; Apollo, the maturity wall, 2 July 2026.
• Fortune, DeepSeek V4 prices and Huawei chips, 24 April 2026.
• Fortune, US utility spending reaches $1.4 trillion, 14 April 2026; Ember, grids for data centres in Europe, 2026.
• Rest of World, the US-Iran war and Gulf AI infrastructure, March 2026; Stimson Center, Iran and Gulf data centres and cables, 2026.
• SQ Magazine, ChatGPT users and subscribers, 2026, citing OpenAI.
• SiliconANGLE, GPU depreciation schedules, 22 November 2025; Data Gravity, how long does a GPU last?, 2026; The Register, AI's circular deal economy, 4 November 2025.
• Meanwhile in Markets, the dark fibre comparison; TechFinitive, telecom failures 2001 to 2003.

Further Reading

• Carlota Perez, Technological Revolutions and Financial Capital (2002): why great technologies arrive with financial bubbles, and what follows.
• Edward Chancellor, Devil Take the Hindmost (1999): a lively history of financial speculation.

Read More

ESCAPING CHIANG MAI HAZE SEASON

Could the best escape from Chiang Mai's smoke be to a city
that has a serious smog problem of its own?

Fast Track

• Part 1 set out the first rule of escape: leave the Lower Mekong airshed, not just the city. This part adds a second: check the destination's own pollution calendar.
• Uzbekistan, Kazakhstan and Kyrgyzstan lie thousands of kilometres outside the burning belt. Their cities do have bad air, but it comes mainly from coal heating in winter and has usually eased by April.
• Uzbekistan is the strongest all-round choice for March and April: mild spring weather, the Silk Road cities of Samarkand, Bukhara and Khiva, and the Navruz festival on 21 March. Kazakhstan is the easiest to reach from Thailand; Kyrgyzstan suits nature lovers, but mostly from April.
• China illustrates both rules at once. Yunnan sits downwind of the same fires as Chiang Mai, and Beijing's spring brings Gobi dust storms. China's 30-day visa-free scheme for many nationalities currently runs only to 31 December 2026.

Introduction

Part 1 ended with a simple consideration: Chiang Mai's March smoke belongs to a regional airshed, so escaping it means leaving that airshed. For most residents that means heading south to the Thai islands or maritime South-East Asia.

But the beach is not for everyone, and March and April are the hottest months on the Andaman coast. Some would rather spend six weeks walking through ancient cities in a cool spring. Central Asia and China are the obvious candidates, yet both have reputations for poor air. This article asks whether those reputations matter in March and April. The answer turns on a second rule: a destination is only a refuge if its own bad season does not coincide with yours.

Why It Matters

For anyone with asthma, a heart condition or young children, choosing the wrong destination can mean swapping one form of bad air for another. Flights, visas and accommodation for a long stay are booked months ahead, before anyone knows how bad the season will be. Annual pollution averages are little help: what matters is the air during the weeks you are actually there.

Contents Cover the Following

1. The second consideration every city has its own pollution calendar
2. Uzbekistan: the Silk Road in spring
3. Kazakhstan: easy to reach, mountains on the doorstep
4. Kyrgyzstan: a nature refuge that wakes up in April
5. China: both rules at work
6. Practicalities: getting there and getting in

1. The second consideration: every city has its own pollution calendar

Central Asia passes Part 1's test easily. Tashkent and Almaty lie well over 3,000 kilometres from Chiang Mai, far beyond the reach of the fires in Shan State and northern Laos.

But leaving one airshed means entering another. Almaty and Bishkek sit on plains hemmed in by mountains, a bowl-shaped setting that, like Chiang Mai's basin, traps pollution. In winter, households burn coal in stoves and both cities rely on Soviet-era coal-fired power stations. On bad days their air rivals Delhi and Lahore, and similar smog affects Tashkent and Dushanbe.

The crucial word is winter. Researchers at Nazarbayev University find Almaty's pollution most severe from October to March, with an average winter PM2.5 concentration of 94 µg/m³, more than six times the World Health Organization's daily guideline. The cause is heating, so the problem eases as the heating season ends. Chiang Mai's calendar runs almost exactly the other way, at its best from November to February and its worst in March and April.

The fit is not perfect. Kazakhstan's weather service, Kazhydromet, issued pollution warnings for Almaty on 26 and 27 March and on 2 and 9 April 2025, when still weather let pollution build up. The fair conclusion is that early March in Almaty can still be smoggy, while April is usually much clearer.

Escaping Chiang Mai's haze is a matter of timing as much as distance: the best refuge is a place whose bad season you can miss.

2. Uzbekistan: the Silk Road in spring

Uzbekistan is the strongest all-round choice, mainly because of its cities. Samarkand, Bukhara and Khiva were wealthy stops on the Silk Road and remain among the most complete Islamic townscapes in the world: turquoise domes, tiled madrasas and old bazaars. Samarkand's Registan and the tombs of Shah-i-Zinda are the best known; Khiva's walled inner town survives almost entire.

The timing suits too. Navruz, the Persian New Year, falls on 21 March and is the region's great spring festival, with music, public celebrations and special dishes such as sumalak, a sweet paste of sprouted wheat. In Samarkand and Tashkent, March highs are typically in the mid-teens Celsius; April, in the low to mid-twenties, is close to ideal for walking.

Tashkent shares the region's winter smog, so early March is the weakest point. Short dust storms can also affect the west, around Khiva and the dried-out Aral Sea. Citizens of around 90 countries can visit visa-free for up to 30 days, and there are nonstop flights from Bangkok to Tashkent. Uzbekistan suits those who want history, architecture and food; those seeking mountains should look further east.

3. Kazakhstan: easy to reach, mountains on the doorstep

Kazakhstan's great advantage is access. Air Astana flies six times a week from Bangkok to Almaty and Thai AirAsia X three times, and Air Astana also flies daily between Almaty and Phuket.

Almaty is a leafy, café-filled city with snow-capped mountains rising directly behind it. The ski slopes of Shymbulak usually still have snow in March, and Charyn Canyon and the Kolsai lakes make good trips once spring arrives. For history, Kazakhstan is thinner than Uzbekistan; the great exception is the UNESCO-listed mausoleum of Khoja Ahmed Yasawi in Turkistan, near Shymkent.

Kazakhstan is the least suited of the three to an early escape. In March, Almaty's highs are typically around 10°C and snow is still possible; April brings highs in the mid to high teens. With the region's most notorious air, its case strengthens as April approaches. It is the most expensive of the three, but also the most comfortable for city life.

4. Kyrgyzstan: a nature refuge that wakes up in April

Kyrgyzstan is the region's outdoor country, with high mountains, the vast lake of Issyk-Kul, yurt camps and a living nomadic culture. It is also the cheapest of the three.

The difficulty is the season. In March and April the high valleys and most trekking routes are still under snow, and highlights such as Song-Kul lake are generally reached from June. A spring visitor is largely limited to Bishkek, the shores of Issyk-Kul and lower day trips such as the Ala-Archa gorge or the medieval Burana Tower. Bishkek shares Almaty's winter smog pattern, and there are no direct flights from Thailand.

Kyrgyzstan works best as the second half of a trip: Uzbekistan's cities in March, then Kyrgyzstan's lakes and foothills in April.

5. China: both rules at work

China is tempting because it is so close. China Eastern flies from Chiang Mai to Kunming, the self-styled "Spring City", in under two hours, and Chengdu Airlines flies to Chengdu three times a week. Yet China illustrates both rules.

5.1 Yunnan: the airshed rule

Yunnan lies downwind of the same fires that fill Chiang Mai's valley. Chinese researchers describe spring smoke from South-East Asia as the decisive obstacle to cleaner air in south-west China. From 31 March to 2 April 2019, PM2.5 in seven Yunnan cities exceeded China's national standard of 75 µg/m³, and modelling presented in 2024 attributed about 60 per cent of Yunnan's burning-related PM2.5 to fires along the Myanmar–Thailand border. Kunming may be cleaner than Chiang Mai at the peak, but it is not outside the airshed, and Xishuangbanna, closest to the border, is the least reliable refuge in China.

5.2 Northern China: the calendar rule

Beijing's winter smog has improved considerably, but March and April are its sandstorm season, when winds carry dust from the Gobi Desert. In March 2021 the city suffered its biggest sandstorm in a decade, and in 2023 it had recorded eight by early April. Storm days are a minority, but this is the wrong season for six weeks of clean air.

5.3 The rest of China and the visa question

Elsewhere, including Shanghai and Hangzhou, neither the fires nor the Gobi dust dominates in spring; check the specific city's recent spring record. China allows citizens of more than 45 countries, including most of the European Union, Australia, New Zealand and, since 17 February 2026, the United Kingdom and Canada, to visit visa-free for up to 30 days. The scheme currently runs to 31 December 2026, and no extension into 2027 has yet been announced for most of these countries, so check before booking. Others, including Indonesians, can use the 240-hour (10-day) visa-free transit with a confirmed onward ticket to a third country.

6. Practicalities: getting there and getting in

Entry rules vary by nationality and change often, so the visa column gives the general position only.

Destination	Getting there from Chiang Mai	Entry (general position, October 2026)	March	April
Uzbekistan	Via Bangkok; nonstop flights to Tashkent	Visa-free for about 90 nationalities, up to 30 days	Cool, some rain; good from mid-March	Ideal
Kazakhstan	Via Bangkok or Phuket; nonstop to Almaty	Visa-free for many nationalities, typically 30 days	Cold; smog possible	Good
Kyrgyzstan	Via Almaty or Tashkent	Visa-free for many nationalities, 30 to 60 days	Cold; mountains closed	Good in the lowlands
China (Yunnan)	Direct to Kunming	30-day visa-free for listed countries, to 31 December 2026	Smoke risk	Smoke risk
China (Beijing)	Via Bangkok or a Chinese hub	As above	Dust storm season	Dust storm season

Tajikistan and Turkmenistan are left out: many nationalities need a visa for Tajikistan, whose mountains are mostly a summer destination, and Turkmenistan requires a letter of invitation.

The most efficient plan follows the two calendars. Leave Chiang Mai in early or mid-March for Uzbekistan, ideally in time for Navruz. Move on in April to Almaty or Kyrgyzstan once the heating season has ended. Return in late April or early May, when Chiang Mai's season usually winds down. Check live air quality readings before booking and again before travelling.

---

Bottom Line

Part 1 answered where the smoke is not; Part 2 adds when. A refuge must lie outside the Lower Mekong airshed, and its own bad season must fall at a different time of year. Central Asia meets both conditions better than its reputation suggests, because its smog is driven by winter heating while Chiang Mai's is driven by spring fires. That makes Uzbekistan the strongest choice for most readers, with Kazakhstan and Kyrgyzstan improving as April arrives. China shows what happens when either condition fails.

The wider lesson is that air pollution is less a property of places than of places at particular times. For anyone choosing where to live, not just where to holiday, the question is not "how clean is the air there?" but "when is it clean, and when am I there?"

Related Articles

• Understand Chiang Mai haze season: where the smoke comes from and where the clean air is (Part 1)

Glossary

Airshed - a geographical area within which air and pollution circulate as a shared system, often bounded by mountains or prevailing winds.

Biomass burning - the burning of vegetation, including forests, crop stubble and leaves, as distinct from burning fossil fuels.

Kazhydromet - Kazakhstan's national weather and environmental monitoring service, which issues air pollution warnings.

Letter of invitation - a document from a host or agency in the destination country, required in support of some visa applications.

Madrasa - a traditional Islamic college, often built around a courtyard with tiled portals.

Navruz - the Persian New Year, celebrated on or around 21 March across Central Asia, Iran and beyond; also spelt Nowruz or Nauryz.

PM2.5 - airborne particles 2.5 micrometres or less across, small enough to penetrate deep into the lungs and enter the bloodstream.

Silk Road - the network of overland trade routes that linked China with Central Asia, the Middle East and the Mediterranean from antiquity to the early modern period.

Visa-free transit - permission to enter a country without a visa for a limited period while travelling on to a third country, usually with a confirmed onward ticket.

References

• Astana Times, "Kazakhstan to Launch Direct Flights From Almaty to Langkawi, Astana to Bangkok", October 2026 - current Kazakhstan–Thailand routes and frequencies.
• Roame, Bangkok to Tashkent route - nonstop service between Bangkok and Tashkent.
• Nazarbayev University Repository, PM2.5 in Almaty - seasonal pattern and average winter PM2.5 in Almaty.
• Prothom Alo / AFP, "Smog veils Central Asia cities" - causes of winter smog in Almaty, Bishkek, Tashkent and Dushanbe.
• Tengrinews, pollution news archive - Kazhydromet air pollution warnings for Almaty in March and April 2025.
• Embassy of Uzbekistan in Malaysia, visa information - visa-free regime for 90 countries, up to 30 days.
• Invest Kyrgyzstan, visa-free regime - Kyrgyzstan's visa-free arrangements by nationality.
• GOV.UK, Tajikistan entry requirements - British citizens require a visa.
• Frontiers in Environmental Science, biomass burning from South-East Asia and air quality in Kunming, 2023 - spring smoke transport into Yunnan; the March–April 2019 episode.
• EGU General Assembly 2024, abstract EGU24-1471 - share of burning-related PM2.5 in Yunnan from the Myanmar–Thailand border.
• Peking University Journal, cross-border transport and air quality in Yunnan border cities, 2025 - Myanmar's dominant influence on Yunnan border cities.
• Deccan Herald, "Thick sandstorms shroud Beijing, several provinces in China" - Beijing's March–April sandstorm season.
• South China Morning Post, eighth sandstorm of 2023
• The Peninsula / Reuters, Beijing's biggest sandstorm in a decade, 15 March 2021
• Skyscanner, Chiang Mai to Kunming - direct China Eastern service.
• DirectFlights, Chiang Mai to Chengdu - Chengdu Airlines, three weekly flights.
• Business Today, China extends visa-free travel for 45 countries to end-2026
• China–Britain Business Council, "China Grants 30-Day Visa-free Travel to British Citizens"
• Interfax, China extends visa-free policy for Russia to 31 December 2027
• Xinhua, China adds Indonesia to 240-hour visa-free transit programme

Further Reading

• IQAir and the World Air Quality Index project - live and historical air quality readings for most of the cities in this article.

Read More

UNDERSTAND CHIANG MAI HAZE SEASON

Understand Chiang Mai's haze season.

Fast Track

• Chiang Mai's haze peaks from early March to late April. In 2026 the city briefly ranked as the world's most polluted major city, while some northern districts recorded more than seven times Thailand's daily limit.
• The smoke does not come mainly from the city. It comes from forest fires and crop burning across a shared regional airshed covering northern Thailand, Myanmar's Shan State and northern Laos.
• Escaping therefore means leaving that airshed, not merely leaving the city. Pai, Chiang Rai and Luang Prabang usually offer no relief; southern Thailand and maritime South-East Asia usually do.
• For those who stay, one properly sealed room with a purifier does more than any mask. A long-term fix depends on a Clean Air Bill still stuck between Thailand's two parliamentary chambers.

Introduction

For four months of the year, Chiang Mai is close to ideal. From November to February the mornings are cool, the skies are clear and Doi Suthep stands sharp above the old city. Then, within a few weeks, the mountain disappears.

By late March 2026, all 111 of Chiang Mai University's monitoring stations in the province showed red. On the morning of 1 April the city recorded an AQI of 235 on IQAir, its fourth consecutive day as the most polluted major city on Earth. Pharmacies reported queues for N95 masks.

The instinctive response is to get out. Yet each year some residents drive to Pai, or take the slow boat to Luang Prabang, and find the same grey sky waiting. Others fly to Krabi and breathe easily within hours. The difference has little to do with how far they travel.

This article asks why. It traces where the smoke actually comes from, explains why the problem is regional rather than local, and uses that explanation to identify where clean air can reliably be found in March and April. It ends with what to do if leaving is not an option, and whether the problem is likely to improve.

Why It Matters

Chiang Mai's annual average for fine particulate pollution looks moderate, but that average conceals a short, severe spike. Around March the monthly PM2.5 concentration rises to many times the World Health Organization's guideline.

The health effects are measurable. A study of upper northern Thailand found that hospital readmissions for chronic obstructive pulmonary disease (COPD) peak in March. Nationally, the Public Health Ministry counted around 3 million people with PM2.5-related illness in 2024.

The haze also shapes the local economy. Tourism thins out, outdoor workers have no choice but to breathe it, and a growing number of residents now plan their year around six weeks away. For anyone with asthma, a heart condition or young children, deciding when and where to go is a genuine health decision, not a lifestyle preference.

Contents Cover the Following

1. A city built in a bowl
2. Where the smoke really comes from
3. The airshed rule: why distance is not enough
4. Where the clean air is in March and April
5. If you stay: turning one room into a refuge
6. Will it get better?

1. A city built in a bowl

Chiang Mai sits in a basin, ringed by mountains. For most of the year that setting is the city's chief attraction. In the dry season it becomes a trap.

The last significant rain usually falls around the end of October. By March the land has been drying for four or five months, and the forests and fields are primed to burn. At the same time, the surrounding mountains limit the movement of air across the valley floor. On still days, a layer of warmer air can settle above the basin like a lid, holding smoke close to the ground until wind or rain breaks it up. The pollution is not evenly spread, either: on the morning of 1 April 2026, PM2.5 readings ranged from about 80 µg/m³ in Suthep, near the city centre, to 286 µg/m³ in Chiang Dao to the north.

Geography does not create the smoke, but it decides how long the smoke stays. That distinction matters, because it explains why the problem is so stubborn, and why a city with modest traffic can briefly outrank Delhi or Lahore.

2. Where the smoke really comes from

The smoke is overwhelmingly rural. Research on northern Thailand estimates that burning vegetation, known as biomass burning, accounts for nearly half of Chiang Mai's PM2.5 across the whole year, and more than 70 per cent of it in April. Vehicles, industry and charcoal stoves add to the load, but they are not what turns the sky white in March.

Three kinds of fire dominate.

Forest fires. These are the largest single source. On 31 March 2024, satellite data from Thailand's space agency, GISTDA, showed 1,477 of the country's 1,864 hotspots in forest and reserved areas, against 294 on farmland. Thai officials have said that around 90 per cent of northern hotspots are started by people, whether to clear undergrowth, gather forest products or for other local reasons.

Crop burning. Farmers burn stubble to prepare fields for the next planting, because it is the cheapest way to clear steep land. Maize is the crop most closely linked to the haze. Greenpeace Southeast Asia estimated that more than 11.8 million rai (about 1.9 million hectares) of forest in northern Thailand, Myanmar's Shan State and northern Laos was converted to maize between 2015 and 2023.

Smoke from across the border. On the same day in March 2024 that Thailand logged 1,864 hotspots, Myanmar logged 5,296 and Laos 3,047. When the wind blows from the west or north, much of that smoke ends up over Chiang Mai.

How much of the haze crosses the border is disputed. A former head of Thailand's Pollution Control Department once put Myanmar's share of northern dust at around 5 per cent, emphasising domestic fires. Researchers at Chiang Mai University and environmental groups place far more weight on transboundary smoke, especially in years when the region's hotspot count surges. The honest conclusion is that the share varies with the wind and the year, and that both domestic and cross-border fires matter.

There is a less visible thread connecting these fires. Thailand is one of the world's largest producers of animal feed, and maize makes up more than a third of it. Cross-border contract farming expanded after Thailand removed tariffs on maize imports from its neighbours. An interpretation follows, though it is not the whole story: part of northern Thailand's haze is a hidden cost of cheap chicken and pork, paid in the lungs of people who live far from the supermarket shelf.

This also explains why local bans struggle. Chiang Mai banned open burning from 1 January to 31 May 2026, yet the province still recorded 10,849 hotspots by 27 April. A provincial ban cannot reach a hillside in Shan State.

The smoke that fills Chiang Mai's valley is not only Chiang Mai's smoke. Escaping it means leaving a region, not a town.

3. The airshed rule: why distance is not enough

An airshed is an area within which air, and therefore pollution, circulates as a shared system. Analysts describe northern Thailand, eastern Myanmar and northern Laos as one such system: the Lower Mekong airshed. Bangkok and the lower central plain form a separate one.

This single idea explains most of the confusing experiences people report. Pai and Chiang Rai lie in valleys within the same airshed, surrounded by the same fires, and are often just as smoky as Chiang Mai, sometimes worse. Northern Laos is frequently worse still: in one recent season, Luang Prabang recorded hazardous readings above 500 on the AQI scale, and travellers routinely describe the old royal capital vanishing behind haze in March.

Bangkok is a less reliable refuge than it appears. The capital has its own smog season, driven by traffic, industry and burning in the central provinces, which usually peaks in January and February rather than March. In 2025 its smog episode ran from early January to late March. By April Bangkok's air is often acceptable, but it is rarely clean.

The practical rule follows directly. The best destinations are those furthest from the interior farmland of mainland South-East Asia and closest to steady sea winds. A short flight south across that boundary achieves far more than a long drive within it.

4. Where the clean air is in March and April

Southern Thailand is the default choice. There is comparatively little agricultural burning, and breezes off the Andaman Sea and the Gulf of Thailand keep the air moving. Further afield, maritime South-East Asia sits outside the mainland burning belt entirely.

Destination	Why the air is usually cleaner	Trade-offs
Krabi and Koh Lanta	Andaman sea breezes; little crop burning nearby	Hot; Koh Lanta is popular with Chiang Mai's remote workers, so book early
Phuket	Island setting; the widest choice of flights and accommodation	Tourist prices; local rubbish burning can still cause short spikes
Koh Samui and Koh Phangan	Gulf breezes; island air	Higher costs on Samui; Phangan is quieter but slower to reach
Koh Chang (Trat)	Coastal and forested, on the eastern Gulf	Long journey; better for budget travellers than short trips
Penang, Malaysia	Outside the Mekong burning belt; easy overland or by air	Warm and occasionally wet; Malaysia can suffer Indonesian haze later in the year, not usually in spring
Bali, Indonesia	Maritime climate; the rainy season is ending by March	Higher costs in popular areas
Cebu, Philippines	February to April are among its driest months	Further away; fewer direct connections from northern Thailand
Nha Trang and Da Nang, Vietnam	Central coast; Nha Trang's best season is February to April	Check readings, as coastal Vietnam is not immune to regional smoke

Three cautions apply wherever you go. First, March and April are the hottest months across the region, and the beaches are hot too. Second, Thai school holidays and the Songkran festival in mid-April push up prices and fill flights. Third, no destination is guaranteed. Local burning happens almost everywhere in South-East Asia, so check live readings for your destination before you book, and again before you travel.

On timing, residents commonly leave for four to eight weeks. Going from about the start of March until late April covers the worst of most seasons. In 2026 the province was down to two active hotspots by 30 April, and officials were already describing the season as winding down.

5. If you stay: turning one room into a refuge

Leaving is not possible for everyone. Work, schooling, caring responsibilities and cost all keep people in the city. For them, the evidence points clearly to one strategy: protect the air you breathe for the most hours, which usually means the bedroom.

Seal one room, then filter it. A purifier in a leaky room struggles to keep up. Close and tape windows, keep the door shut, set the air conditioner to recirculate, and run a HEPA purifier sized for the room. A home monitor will tell you whether it is working: if the indoor reading creeps towards the outdoor one, air is getting in.

Consider positive pressure. A positive-pressure room adds a fan that pushes filtered outside air into the sealed room, so air leaks outwards through the gaps rather than smoke leaking in. Hospitals and child development centres across the north use the approach. The low-cost clean-air room developed at Chiang Mai University, now being rolled out to 83 sites in eight northern provinces, costs about THB3,600 per room.

Wear a mask that seals. An N95, KN95 or FFP2 mask that fits snugly over the nose and cheeks filters fine particles. A surgical mask does not. Buy them before the season: during the late-March spike in 2026, pharmacy sales of N95 masks rose by about 30 per cent.

Check the air every morning, at the nearest station. Readings vary sharply by district and by day. On 23 March 2026 Suthep recorded just 15 µg/m³ under a clear sky; within a week the city topped the global rankings. IQAir is the easiest app to use; the government's Air4Thai network is the official record but reports 24-hour averages that lag behind sudden spikes; Chiang Mai University's DustBoy sensors give real-time readings by neighbourhood; and GISTDA's fire map shows where fires are burning.

The principle behind all four steps is the same as the airshed rule, applied at the scale of a home. You cannot clean the valley, but you can create a small, separate pocket of air within it.

6. Will it get better?

Policy is moving, but slowly. Thailand tightened its national PM2.5 standard in 2023, lowering the 24-hour limit from 50 to 37.5 µg/m³. That is still two and a half times the World Health Organization's guideline of 15 µg/m³.

The more significant reform is the Clean Air Management Bill, which grew out of a citizen-led campaign and would recognise a right to clean air, create national and provincial clean-air committees and introduce a polluter-pays system. Its passage has been tortuous:

• October 2025: the House of Representatives passed it by 309 votes to none.
• December 2025: parliament was dissolved ahead of a general election, leaving the bill in limbo.
• May 2026: after the election, parliament voted 611 to 3 to revive it.
• July 2026: the Senate returned it with amendments, including greater business representation on clean-air committees and safeguards against duplicate fees.
• September 2026: the House rejected the Senate's changes by 414 votes to 2, sending the bill to a 20-member joint committee.

At the time of writing, the bill is not yet law. Critics of the Senate amendments, including the People's Party MP for Chiang Mai who led the House rebuttal, argue that the revisions favour big business and weaken local control. Senators maintain that they support the bill in principle but disagree on specific provisions.

Even a strong law would face the limit identified in Section 2. Much of the smoke originates outside Thailand, in areas where enforcement is weak or, in parts of Myanmar, where the state has little authority at all. Regional agreements exist: in 2020 the five Lower Mekong governments adopted the Chiang Rai Plan of Action, aiming to cut hotspots by 40 per cent by 2025. The severity of the 2026 season suggests that such targets have not yet translated into clean skies.

The most plausible near-term outlook is therefore gradual improvement at best, with large swings from year to year depending on rainfall and wind. For the next few seasons at least, planning to leave, or to seal a room, remains the sensible default.

---

Bottom Line

Chiang Mai's haze is best understood not as a city's pollution problem but as a region's land-use problem, concentrated by geography into one valley. That single shift in perspective answers the practical question. To escape the smoke, leave the Lower Mekong airshed: go south to the Thai coast or out to maritime South-East Asia, and treat nearby valleys as part of the same system. If you cannot leave, recreate that separation at home with one sealed, filtered room.

The wider implication reaches beyond travel planning. Air does not respect provincial boundaries or national borders, and neither do the markets that drive the fires. A Clean Air Act, however well drafted, can govern only one side of the hills. The question that remains is whether the governments and companies that share the airshed, and profit from what grows in it, will ever treat its air as a shared responsibility.

Deep Dive: Reading the Numbers

AQI is not the same as µg/m³. Most apps display an Air Quality Index (AQI), a scale built from pollutant concentrations, in which 0–50 is good, 151–200 unhealthy and 201–300 very unhealthy. Thai government figures and health guidelines use the PM2.5 concentration itself, measured in micrograms per cubic metre (µg/m³). An AQI of 235 and a PM2.5 reading of 235 µg/m³ are very different things, so check which one you are looking at.

Averages hide spikes. Thailand's official Air4Thai readings are 24-hour averages. During a sudden worsening they read lower than real-time sensors such as DustBoy. For deciding whether to go for a run this morning, the real-time figure is more useful; for comparing days, the average is fairer.

The standards differ. Thailand's 24-hour limit has been 37.5 µg/m³ since June 2023, matching the WHO's third interim target. The WHO's actual guideline is 15 µg/m³. A day that is "within the Thai standard" may still be more than double the WHO guideline.

The cigarette comparison. Berkeley Earth offers a rough rule of thumb: breathing air at 22 µg/m³ of PM2.5 for a full day is broadly comparable, in health terms, to smoking one cigarette. By that measure, central Chiang Mai's air on 1 April 2026 was equivalent to around four or five cigarettes a day, and Chiang Dao's to about thirteen. The comparison is an approximation, but it conveys scale more vividly than any index.

Glossary

AQI (Air Quality Index) - a scale that converts pollutant concentrations into a single number, usually from 0 to 500, with colour bands from "good" to "hazardous".

Airshed - a geographical area within which air and pollution circulate as a shared system, often bounded by mountains or prevailing winds.

Biomass burning - the burning of vegetation, including forests, crop stubble and leaves, as distinct from burning fossil fuels.

COPD (chronic obstructive pulmonary disease) - a long-term lung condition, including emphysema and chronic bronchitis, that air pollution can worsen.

HEPA filter - a high-efficiency particulate air filter, designed to capture the very fine particles that make up PM2.5.

Hotspot - a fire, or an area of unusually high heat, detected by satellite; used as a measure of how much burning is taking place.

N95, KN95 and FFP2 - US, Chinese and European standards for respirator masks that filter at least 94–95 per cent of fine airborne particles when properly fitted.

PM2.5 - airborne particles 2.5 micrometres or less across, small enough to penetrate deep into the lungs and enter the bloodstream.

Positive-pressure room - a sealed room into which a fan pushes filtered outside air, so that air leaks out through gaps rather than polluted air leaking in.

Rai - a Thai unit of land area equal to 1,600 square metres; 6.25 rai make one hectare.

Temperature inversion - a weather condition in which a layer of warmer air sits above cooler air near the ground, trapping pollution beneath it.

Transboundary haze - smoke pollution that originates in one country and drifts into another.

References

• Thailand Starter Kit, "Chiang Mai Burning Season for Expats", October 2026 - 2026 season timeline, station readings for 1 April 2026, GISTDA hotspot counts for 31 March 2024, clean-air room costs and mask sales.
• Thai Rath, "Chiang Mai Imposes Five-Month Ban on Open Burning", 2 January 2026
• Thai Rath, clear skies over Chiang Mai, 23 March 2026
• Thai Rath, Chiang Mai ranked worst for air quality, 27 March 2026
• Khaosod English, "Chiang Mai air quality improves as hotspots decline", 30 April 2026
• NASA Earth Observatory, "Smoke Shrouds Northern Thailand"
• Near real-time biomass burning PM2.5 emission estimation in northern Thailand (PMC)
• Decadal trends in air pollutants and COPD in upper northern Thailand, 2013–2022 (PMC)
• The Nation, PM2.5-linked illness figures, 1 April 2026
• Thai PBS World, "90 percent of hotspots in northern Thailand caused by human activity"
• Greenpeace Southeast Asia, "Maize, Land Use Change and Transboundary Haze Pollution"
• Radio Free Asia, "Forest encroachment for maize cultivation increased in Lower Mekong", September 2023
• LiCAS News, maize and forest loss in the Lower Mekong, September 2023
• Bangkok Post, "New PM2.5 standard will alarm public"
• Mekong Eye, haze in Luang Prabang
• Wikipedia, "2025 Bangkok smog"
• Office of Natural Resources and Environmental Policy and Planning, revised PM2.5 standard (Thai)
• Nation Thailand, "Thai MPs reject clean air changes", 3 September 2026
• UCA News / AFP, "Thailand advances clean air bill targeting polluters", 15 May 2026
• Lowy Institute, The Interpreter, "Hazy future: Thailand's Clean Air Bill"
• Thai Rath, "Senators Consider Rejecting Clean Air Bill, Admit Some Disagreements", 6 May 2026
• Berkeley Earth, air pollution and cigarette equivalence
• Tieland to Thailand, "10 Places to Escape to During Northern Thailand's Burning Season" (2018)

Further Reading

• Mekong Eye, "Transboundary haze: the hidden price of cheap maize" - an accessible account of the maize supply chain and the farmers caught in it.
• NASA, "Grappling with Thailand's Seasonal Haze" - the 2024 ASIA-AQ airborne campaign over Chiang Mai and Bangkok.
• Nation Thailand, "Thailand faces mounting climate risks amid Clean Air Bill delay", September 2026 - the campaigners' perspective on the bill's long journey.
• Chiang Mai University Climate Change Data Center (DustBoy) and GISTDA fire map - live tools worth bookmarking before the season.

Read More

LA GESTION DU DÉSORDRE : LES ÉMEUTES EN FRANCE, OCTOBRE 2026

Comment une fuite dans le toit d'un lycée de banlieue finit-elle, quatre semaines plus tard, par une mairie en flammes à Limoges ?

En bref

Les émeutes du 17 octobre 2026 n'ont été plannifiées d'avance par personne, mais elles ne sont pas tombées du ciel. Comme les soulèvements français précédents, elles suivent six dynamiques récurrentes : l'accumulation des tensions, l'étincelle, la contagion, l'escalade, la diffusion, puis la tentative de rétablir l'ordre. L'État retrouvera sans doute le contrôle de la rue. Rien ne garantit qu'il retrouvera la stabilité.

1. Un pays qui se soulève régulièrement

En France, la contestation de la jeunesse a souvent payé : retrait de la loi Devaquet en 1986, abandon du CPE en 2006. Quand la colère venait des banlieues sans revendication unique, comme en 2005 et en 2023, l'ordre a été rétabli par la force et les griefs sont restés intacts. Quand une revendication précise s'est muée en révolte générale, comme en 1968 et avec les Gilets jaunes, l'État a mêlé concessions et répression, et la colère a fini dans les urnes. 2026 combine les trois cas de figure, ce qui la rend difficile à éteindre.

Et plus largement, on comprend la colère. Les prélèvements obligatoires atteignent les niveaux les plus élevés d'Europe, et la dépense publique absorbe la plus grosse part du PIB de toute l'Union européenne (environ 58 % actuellement !). Pourtant, les services publics se sont dégradés au point que ces revendications légitimes restent sans écoute. Résultat : un État démesuré par ce qu'il coûte, mais rétréci dans ce qu'il rend. Il prétend tout prendre en charge (« l'État, c'est moi »), mais faillit à sa mission première : assurer les services sociaux et l'ordre public.

Que doivent ressentir les étudiants ? Ils ne parviennent pas à s'imaginer un avenir dans un système qui semble vivre ses dernières convulsions. Ce qu'ils voient, en revanche, ce sont les dettes qu'on attend d'eux qu'ils remboursent : des dettes contractées pour des projets absurdes et corrompus, sur lesquels on ne les a jamais consultés, lancés par une élite lointaine de responsables politiques intéressés, issus des grandes écoles (dont Macron, soit dit en passant, est l'archétype).

2. L'accumulation : ce qui couvait avant le premier incident

Près d'une heure de cours sur dix n'est pas assurée dans le secondaire public. Les classes comptent environ 25 élèves au collège, parfois près de 40 au lycée, alors que 30 est un plafond raisonnable. Les travaux de Piketty et Valdenaire montrent que réduire les effectifs profite d'abord aux élèves défavorisés. Surtout, l'école ne mène plus nulle part : 21,1 % des 15-24 ans actifs sont au chômage.

La tension est la plus forte en Seine-Saint-Denis et dans le Val-de-Marne, où le mouvement est né. C'est aussi là que se concentre la population immigrée (environ 32 % en Seine-Saint-Denis, contre 10 % en moyenne nationale). Le lien est géographique et social, pas ethnique : les lycées les plus dégradés sont dans les quartiers les plus pauvres, et leurs élèves, en majorité français, ont le moins à perdre.

S'y ajoutent l'austérité (54 milliards d'euros d'économies prévues pour 2027), une instabilité politique inédite (quatre Premiers ministres depuis début 2024) et le choc pétrolier de la guerre en Iran. Le nucléaire a protégé les factures d'électricité ; la hausse est passée par l'essence et l'alimentation. Dans un budget gelé, une seule ligne augmente : la défense.

3. L'étincelle : un lycée de Créteil

Le 17 septembre, les élèves du lycée Saint-Exupéry à Créteil se mobilisent contre le manque de professeurs et des bâtiments délabrés. Des revendications que presque tous les parents peuvent soutenir. Elles trouvent un terrain prêt : les canaux Telegram du mouvement « Bloquons tout » de 2025, endormis mais intacts. Une date, le samedi 17 octobre, circulait déjà sur TikTok depuis le 8 septembre, sans que personne ne la pilote.

4. Contagion, escalade, diffusion

Fin septembre, les blocages touchent environ 400 lycées. Début octobre, on compte plus de 6 000 interpellations, des gaz lacrymogènes devant les établissements, et un budget qui ne protège que la défense, présenté en pleine crise. Le 17 octobre, syndicats, Gilets jaunes et retraités rejoignent les lycéens : environ 550 000 manifestants. La mairie de Limoges brûle, celle de Marseille est envahie, et le centre de Toulouse est pillé, avec entre 32 000 et 80 000 manifestants et 1 500 à 2 000 casseurs selon le préfet. Les violences sont le fait de quelques milliers de personnes, black blocs et pilleurs, et non du demi-million de manifestants.

5. Le rétablissement de l'ordre, et ce qu'il laisse derrière lui

Le 18 octobre, Emmanuel Macron a promis la fermeté et la transparence sur le maintien de l'ordre, et annoncé une concertation avec lycéens, enseignants et syndicats. Les syndicats n'y voient « rien de concret ». Sans majorité, tenu par ses économies et à six mois de la présidentielle, le gouvernement a peu de marge. L'histoire suggère que les griefs non résolus ne disparaissent pas : ils passent dans les urnes, ressurgissent plus tard, ou nourrissent de nouveaux mouvements.

6. Ce que ces émeutes ne sont pas

Ni une simple affaire scolaire, ni l'œuvre de quelques casseurs. Ni une « révolution de couleur » orchestrée par le Kremlin : rien ne l'indique, et Viginum n'a relevé qu'une ingérence étrangère minime. Ni une conséquence directe de l'immigration : celle-ci explique où la tension était la plus forte, pas pourquoi le pays tout entier a manifesté.

7. La guerre en toile de fond

Les banderoles parlaient d'écoles et de prix, pas de la Russie. À mon sens, la plupart des manifestants se désintéressent de cette guerre, présentée selon eux à coups de propagande par une élite mondialisée lointaine, achetée par des promesses de carrière et de prospérité, et payée avec leurs écoles et leur avenir. Ce n'est pas la guerre qui a provoqué les émeutes, mais le budget qui la finance.

Récap

En conclusion

L'État rétablira l'ordre. Mais une école qui ne mène plus à l'emploi, des banlieues qui se sentent abandonnées et une génération sceptique face au réarmement seront contenues ou déplacées, pas résolues. La vraie question n'est pas de savoir si la France se soulèvera de nouveau, mais lesquelles de ces tensions s'accumuleront d'ici là.

Sources principales

• Élysée : Adresse aux Français, 18 octobre 2026
• France 24 : ce qu'il faut retenir de l'allocution d'Emmanuel Macron
• Politis : les quartiers populaires à l'avant-garde de la mobilisation lycéenne
• Le JDD : « Bloquons tout », qui se cache derrière la mobilisation du 17 octobre ?
• France 3 Occitanie : le 17 octobre à Toulouse
• Insee : localisation des immigrés et des descendants d'immigrés
• Europe 1 : chômage à 8,1 %, 21,1 % chez les jeunes
• Piketty et Valdenaire : l'impact de la taille des classes sur la réussite scolaire (PDF)
• Clubic : la hausse de l'électricité en août sans lien avec l'Iran
• Europeans Today : les émeutes et la crainte d'une ingérence russe (en anglais)

Version intégrale en anglais, avec toutes les références :

Read More