2027: Risks, Opportunities and Kunkafa’s Forecasts, From AI to Popular Unrest

At Kunkafa we read every market on every duration, every morning. A new year is not a date for us. It is a set of distances: how far each market could travel, which way, and how confident Kunkafa is about each path.

Those distances are read from the markets’ own data. But the world behind the data is changing, and going into 2027 there are more risks and major challenges than at any time since 2008.

This analysis takes each challenge in turn, with data we could verify, and ends with the opportunities they open up.

What this report covers:

• AI keeps getting cheaper: prices fall about 13 times a year. So far use has grown even faster, but the profit has not arrived.
• Who pays the real price of AI: most users are subsidised, and pricing could move from tokens to the value an answer creates.
• Open-source AI: open models cost nothing; the memory they need has more than doubled in price.
• Chips without buildings: the largest buyers’ unfinished assets jumped in 2026, and our hypothesis is that Nvidia’s growth slows more than expected.
• Where the data centres could go: US permits are closing; Egypt and Morocco could host, if they bring their own power at night.
• Europe: inflation at 4%, France’s widest borrowing gap since 2012, and protests in the streets.
• China: exports booming on AI hardware, the home economy weak; our reading is growth closer to 4% than 5% in 2027.
• Rates: the cuts did not come. A recession, if it comes, is more likely to arrive through credit than through jobs.
• Two wars, three choke points: fuel, fertiliser and grain are all squeezed at once.
• What people buy: car payments, freight and health premiums squeeze households.
• The subscription economy: a repricing, not a collapse.
• Opportunities and the conclusion: where the pressure creates demand, and how it all fits together.

1. AI keeps getting cheaper. Who keeps the money?

The price of a given level of AI performance has fallen about 47% every quarter for three years. That is roughly thirteen times cheaper each year, according to Epoch AI (Epoch AI).

The labs’ price lists show it. Anthropic’s top model cost $15 per million input tokens in May 2025; today it costs $4. Its smallest went from $1 to $0.10 (Anthropic pricing).

Cheaper should mean less revenue per answer. So far it has meant the opposite.

Anthropic’s quarterly revenue went from $2.04bn at the end of 2025 to $16.96bn in the September quarter. The compute it spends to earn each dollar fell from 144 cents to 58 (our reading of the S-1/A).

When something useful gets cheap enough, people use far more of it. Economists call it the Jevons paradox. Coal, electricity and bandwidth all followed it.

Two things keep this from being the whole story.

• The margin has not arrived. Anthropic’s operating margin was +8% in the June quarter and −3% in September. OpenAI told investors its annualised revenue was close to $50bn in September, and the Wall Street Journal reported a $12.3bn operating loss on $6.7bn of revenue in the June quarter. Revenue grows; profit is still a promise.
• Prices per token fall faster than prices per task. Reasoning models spend many more tokens on each answer. A customer who pays 70% less per token but uses five times as many tokens per job pays more, which is good for the lab until the customer notices.

So 2027 could go either way:

• The squeeze. Open models set a floor near zero for routine work (Chinese open models are offered at a few cents per million tokens), the lead of the frontier labs narrows from months to weeks, and prices fall faster than use grows. Revenue growth slows while the compute bills signed in 2025 and 2026 come due.
• The flood. Cheaper answers pull in work that was never worth automating: every invoice, every call, every line of code. Use grows faster than prices fall, as it did at Anthropic this year, and the labs grow into their commitments.

The number to watch in 2027 is not revenue. It is compute cost as a share of revenue. If it keeps falling while prices fall, the flood is winning. If it turns back up, as it did between the June and September quarters, the squeeze is. We set out the money side in Is AI a bubble?.

2. Who pays the real price of AI?

Most people who use AI do not pay what it costs. ChatGPT had more than 900 million weekly users early this year, and about 50 million paid: roughly one in twenty.

Paid plans are priced like streaming, about $20 a month. The work behind them is priced like electricity: the heavier the use, the bigger the bill.

OpenAI’s bill for running its models was projected to rise from $8.4bn in 2025 to $14.1bn in 2026. In the June quarter it lost $12.3bn from operations on $6.7bn of revenue.

Anthropic shows the other model. Most of its revenue comes from businesses paying per token, closer to the real cost of each answer. That is why its compute bill fell to 58 cents per dollar of revenue.

A flat fee makes light users pay for heavy ones. Investors cover the rest.

That subsidy is a hidden risk for 2027. If investors stop paying for growth, consumer AI plans have two ways to go: up in price, or down in what they include. Caps, slower answers and ads all work like a price rise.

The question is who stays. Light users could drop to the free tier; section 11 shows how ready households are to cancel. Businesses that built their work around AI would pay more rather than unwind it.

From tokens to value. The deeper change could be in what a token costs. Today it costs the same whether it fixes a laptop or helps design a drug.

Anthropic’s chief executive, Dario Amodei, made the point on the Dwarkesh Podcast in February. Telling someone to restart their computer is worth a few cents. Advice to a drug company on one ring of a molecule is not.

“Those tokens could be worth tens of millions of dollars,” he said. He expects business models that charge for results, or by the hour like labour (Dwarkesh Podcast, 13 February 2026).

Customer service already works this way at some vendors, which charge per problem solved rather than per seat or per token. If the labs follow, 2027 could split AI into two markets:

Two markets for AI in 2027: commodity tokens and value tokens

Source: Kunkafa’s analysis, drawing on Dario Amodei’s remarks on the Dwarkesh Podcast, 13 February 2026.

The effects would spread in both directions.

• Demand: frontier models in pharma, finance and engineering would be priced by what they save, which could lift lab revenue far above what tokens suggest.
• Supply: when compute is short, a lab earning a share of a drug’s value would give that customer its best chips first. Ordinary users would feel it as slower service at peak times.
• Industries: they would choose between sharing the value, building on open models, or partnering with a lab for equity instead of fees.

The limit is the open model: a lab can only charge for value a free model cannot deliver. That is why section 3 matters to the labs as much as to the hardware makers.

3. Open-source AI: free to download, very expensive to run

China’s open models are often described as AI’s great equaliser: frontier-class weights anyone can download. The weights are free. The machine they need is not.

Free to download, but the best open models need data-centre hardware: what it takes to run them

Sources: model cards and hosting guides for each model, November 2026; sizes are approximate and change with compression.

The pattern is steady: each year a smaller model matches what a large one did the year before, and compression roughly quarters the memory a model needs. What runs on a laptop in 2027 could match the best paid models of 2025. The frontier, though, keeps moving, and the best models of 2027 are likely to stay in data centres.

The bottleneck is memory, and memory is going the wrong way. Chipmakers are shifting to the high-bandwidth memory AI servers use, which takes about three times the wafer area per gigabyte.

PC memory contract prices rose 105–110% in the first quarter of 2026 alone, then another 58–63% in the second, according to TrendForce. SK Hynix says its 2026 output is sold. Micron closed its consumer brand, Crucial.

A 32 GB memory kit that cost about $72 in mid-2025 sold for $390–560 this autumn. IDC expects the average PC to cost up to 8% more this year.

That leaves an uncomfortable answer to the question of whether super-powerful models could run locally in 2027: yes for good-enough models, no for the best, and at a hardware price that has moved against the user.

• Memory makers (SK Hynix, Samsung, Micron, SanDisk) hold the pricing power for as long as supply stays short. Memory has always been a cycle; new factories ordered at today’s prices tend to arrive just as demand cools.
• PC, phone and console makers pay more for every device and pass it on. Sony, Microsoft and Nintendo have raised console prices this year.
• The labs lose the cheap end of the market to open models but keep the top end, where the hardware a single company cannot buy is the moat.
• Users get capable free models, and pay for them in hardware.

4. They bought the chips. Where are the buildings and the power?

The hypothesis is simple. In 2025 and 2026 the biggest buyers of Nvidia chips raced to secure them, because whoever had chips could serve the demand.

But chips arrive in weeks, and a powered data centre takes years. If buyers ran ahead of their buildings, a growing share of what they bought should be sitting idle.

We tested it three ways: Nvidia’s sales against capacity switched on, the buyers’ own accounts, and the second-hand market.

Test 1: chips shipped against capacity switched on.

• What Nvidia sold. Data-centre revenue was $193.7bn in the year to January 2026, and $164.2bn in the following six months alone ($75.2bn and $89.0bn), from Nvidia’s quarterly releases (SEC filings).
• What that buys in power. Jensen Huang puts Nvidia’s equipment in one gigawatt of AI capacity at about $35bn, and over $40bn with the newest systems. So Nvidia shipped enough for roughly 5 to 5.5 gigawatts last year, and is now shipping at 8 to 9 a year.
• What was switched on. The US, Nvidia’s largest market, added about 8.5 gigawatts in 2025, all data-centre types included (Goldman Sachs). For 2026, 13.6 are scheduled, but only 7 to 8 are likely on time. Of 12 gigawatts of AI capacity announced this year, about 5 were under construction (Goldman Sachs).

In 2025 there was no visible gap: the US alone switched on more capacity than Nvidia’s chips worldwide needed. In 2026 a gap opens. Each gigawatt that waits holds $35–40bn of equipment, so a gap of two or three gigawatts would mean $70–120bn of chips bought and not yet earning.

Test 2: the buyers’ own books. Industry arithmetic can hide a lot, so we read the latest filings of the biggest buyers.

Each reports what it has paid for that is not yet working: “construction in progress” or “assets not yet in service”. Alphabet says these can take “months to years” to switch on. If buyers run ahead of their buildings, that line should grow faster than anything else.

The biggest buyers doubled their spending, and a trillion dollars of leases has not started

Sources: each company’s latest 10-Q and its 10-Q or 10-K a year earlier, as filed with the SEC. Share = the not-yet-in-service line divided by all property and equipment at cost. Microsoft does not report construction in progress.

Company by company, the books say different things:

• Alphabet and Meta: the clearest signal. Their share of property not yet working rose 3 and 5 points through 2025, then another 7 and 6 in nine months of 2026. Together they added $106bn of idle assets this year, twice all of 2025.
• Amazon: rising, then the warehouses. Its construction in progress grew from 8.9% of property in 2023 to 13.4% in 2025. Business Insider reported that AWS takes Nvidia racks before its sites are ready, some stored in warehouses for lack of power.
• Microsoft: the bottleneck in its own words. It does not report the line. But it said September-quarter spending came in below plan because data-centre partners failed to deliver capacity on time.
• The specialists are switching on. At CoreWeave the share not yet working fell from 27.6% to 22.6% in the first half of 2026, and at Oracle it held near a third. The neoclouds, whose business is renting out every chip they own, appear to be energising faster than the giants.

The last column is the bigger story. A year ago, the four largest buyers had signed about $282bn of data-centre leases not yet started. Today it is about $1,016bn, starting between 2026 and 2034.

They did not only race for chips. They raced for buildings and power, years ahead. A trillion dollars of rent is promised for capacity that does not exist yet.

Test 3: the second-hand market. If buyers were stuck with chips, some would sell, and rental prices would fall first. They have not.

An Nvidia H100 rents for about $3.25–4.20 an hour this autumn, from $1.25 on spot listings. Used H100s sell for about $25,000–35,000. There is no sign of chips being dumped.

The verdict. The hypothesis holds, in a precise form. The biggest buyers bought ahead of their buildings, and the share of what they own that is idle rose sharply in 2026.

It does not hold in its loud form. Nobody is throwing chips away, and the specialist clouds are switching theirs on.

Why waiting is expensive. A chip that waits a year for power loses a year against the next generation. Nvidia ships a new design about every year, and buyers write servers off over about six.

The pressure shows at Nvidia too. From January to July, the money customers owe it rose 64%, to $63.1bn, while quarterly data-centre sales rose 43%. Five customers account for 70% of that debt.

Nvidia has also committed $36bn to rent back cloud capacity from the AI clouds it sells chips to (Nvidia 10-Q). Its orders with suppliers, mostly memory makers, rose from $119bn to $279bn in a quarter.

Our hypothesis for 2027: Nvidia’s growth slows more than expected. Nvidia sold $215.9bn in its last full year, up 65%. Analysts expect about $412bn this year and about $692bn next year, growth of about 68%, and Nvidia itself has guided to about 70% (Nvidia, CNBC, consensus via Yahoo Finance). Our reading of the numbers above is that the risk sits below that line:

• The buyers have paid ahead. At Alphabet and Meta more than a quarter of everything they own is not yet working, and the share rose by six to seven points this year. Buyers in that position tend to slow their orders until the buildings catch up.
• The buildings arrive over years, not quarters. A trillion dollars of leases starts between 2026 and 2034. Chip orders that run ahead of buildings have to slow to their pace sooner or later.
• Power and permits are tightening, as section 5 shows, in the states where most of the capacity was planned.
• Customers owe Nvidia more, and owe it faster than its sales grow.

This is a hypothesis built on public numbers, not a forecast of Nvidia’s results. It could be wrong.

Three things would prove it wrong: the buyers’ idle share starting to fall, Nvidia guiding at or above the roughly $124bn expected for the January quarter, and its receivables growing back in line with sales.

The second pressure: questions at the top. Nvidia enters 2027 under more scrutiny than at any point in its history, and none of it has been proven.

• The chief executive. On 8 November Jensen Huang told staff he had had a relationship with a Chinese businesswoman whom US intelligence reportedly believes spied for Beijing. The board hired a former acting US attorney to investigate (CNBC, NBC News).
• The chips claim. The Information reported that her network helped move about 240,000 Nvidia chips, worth at least $7.5bn, to China. China calls the claims groundless, and none of it has been tested in court.
• The smuggling probes. The New York Times and Bloomberg reported that a Commerce Department inquiry into chips moved through a Singapore data centre was transferred to the Justice Department after Nvidia’s general counsel contacted the department (Bloomberg). A separate case concerns about $2.5bn of servers with Nvidia chips (Fortune).
• Competition. The Justice Department is examining Nvidia’s $12.9bn purchase of Hugging Face, alongside a long-running review of its conduct (CNBC).
• Congress. Huang declined to testify before the Senate Banking Committee in June. From January Democrats chair the committees, with the power to subpoena on chip exports and on the financing deals in which Nvidia funds its own customers; senators of both parties have asked for answers (CNBC).

Nvidia’s shares fell 4.3% on 9 November, about $200bn in a day. Scrutiny does not cut orders by itself. Bank of America told clients it “would not equate allegations with proven business damage”.

But the routes to sales are real: tighter export rules, more cautious government buyers, a distracted leadership, and hearings that put the circular financing on the record.

In 2027 Nvidia could face its first slowdown and its first sustained scrutiny at the same time.

If buyers slow their orders to let the buildings catch up, the pause travels down the chain: to Nvidia, to the memory makers, and to the lenders who financed the racks.

If they keep ordering, the gap grows on their balance sheets. The debt side is in our AI bubble analysis. The next test is Nvidia’s results on 17 November.

5. Where the data centres could go instead

The power is not the only thing missing. Permission is.

• New York paused state permits for data centres over 50 MW in July. Oregon paused permits until July 2027. Pennsylvania now requires local approval for anything over 25 MW and makes the operators pay for their own power and lines. Moody’s counts at least 14 states weighing moratoriums.
• In a Quinnipiac poll in September, nearly three in four Americans opposed a data centre in their own community. On 3 November voters backed restrictions in 40 of 41 local ballot measures (Forbes).
• About 40% of the data-centre capacity planned in the US sits in states whose newly elected governors promised to stop or pause construction, according to Cleanview data (Spokesman-Review). We covered the election in the blue wave and AI stocks.

Read next

Blue Wave and AI Stocks: What the Midterms Mean

Could the blue wave sink AI stocks? SOXL, CoreWeave, Micron, SK Hynix and SoftBank are far below their highs. We read the debt and the data-centre plans.

Demand does not go away when a permit is refused. It moves to where there is land, power and permission, as long as the power stays on around the clock.

That opens a door for new countries. Egypt and Morocco are two of the strongest candidates, and neither is a simple story.

Egypt, Morocco or the Gulf: what each offers a data centre, and its main risk

Sources: Egyptian electricity and petroleum ministries via EnterpriseAM and Ahram, CSIS, Reuters, World Nuclear News; Kunkafa’s comparison, November 2026.

Egypt’s case is better than its reputation, and harder than its sunshine. The sun is real. Benban, near Aswan, is one of the world’s largest solar parks, at about 1,465 MW and growing to 1,735 MW.

The government ordered 3,000 MW of solar and 600 MW of batteries before this summer. A 3,000 MW power link with Saudi Arabia is due to start this month (Arab News).

Egypt kept the lights on through summer 2026, after daily cuts in 2023 and 2024. Its cables carry about 17% of the world’s internet traffic and over 90% of Europe–Asia traffic (CSIS).

The hard part is the night. A data centre uses the same power at 3 a.m. as at noon. A campus running on the sun needs three to four times its load in panels, plus batteries for the dark.

This summer Egypt kept its grid running by importing 7.94 million tonnes of LNG in seven months, against 3.4 million a year earlier, part of it from Qatar (EnterpriseAM).

A grid balanced by imported gas cannot also promise spare gigawatts to foreign data centres.

Egypt’s answer to the night is under construction. The El Dabaa nuclear plant, four Russian-built 1.2 GW reactors, is due to connect its first unit in 2028 and all four by 2030 (World Nuclear News, TASS).

That is 4.8 GW running day and night, about an eighth of this summer’s peak. It changes the case for the end of the decade, not for 2027.

Nuclear projects tend to run late, and this one already slipped from 2026. It also ties fuel and service to Rosatom, a Russian state company, which some Western tenants would weigh.

Until then, the winning model is a data centre that brings its own power: its own solar and batteries in the desert, the grid as backup, and the Saudi link for balance.

That is the condition to watch for in every announcement. A 20 MW first phase is a start; a gigawatt campus with its own power would be a shift.

Morocco has the cleaner energy story and the shorter distance to Europe, with less grid at stake. Its 500 MW Dakhla project has no published cost or date, and Dakhla sits in Western Sahara, whose status is disputed, a risk any foreign tenant would price. The 36 MW AI factory backed by Nvidia partners is the nearer test.

The Gulf was the natural host before February. It has money, power and land. Iranian missiles over Gulf cities changed how a twenty-year data-centre lease is priced, at least until the war ends.

6. Europe: prices up, patience down

Europe went into 2026 with inflation under control. It ends the year at 4%.

Energy did it. Energy prices are up 20.8% in a year, while core inflation, which leaves out energy and food, holds at 2.5% (Euronews, citing Eurostat).

Gas passed €100 per megawatt hour in late October, nearly three times February’s price. The European Commission puts the war’s cost to Europe at about €100bn. The ECB raised its deposit rate to 2.50%, and markets lean towards another rise in December.

The money to soften it is thin. Euro-area government debt is 88.8% of GDP. France pays 4.79% to borrow for ten years, 153 basis points more than Germany, the widest gap since the euro crisis of 2012, while its parliament fights over a budget.

And patience is running out. In October hundreds of thousands marched in France, about 550,000 on one day by one count. Truckers blocked roads around Naples and the port of Messina for five days.

Protests spread to Belgium, Portugal, Spain and Albania, and Irish fuel blockades began in April (Al Jazeera, France 24). France elects a president in April 2027, with fuel prices at the centre of the campaign.

Could it become 2011 again? After 2008, unrest came in two waves. In Europe it followed austerity: Greece, and Spain’s indignados. In North Africa and the Middle East it followed food.

Researchers found that food riots clustered when the FAO food price index crossed a threshold, and that the Arab Spring broke out at a peak (Lagi, Bertrand and Bar-Yam, 2011).

Food today is close to the level of 2011, though not rising the way it was then: the index fell 1.2% in October and cereals are 12.6% cheaper than a year ago. The pressure now is energy, and energy takes longer to reach the street. The two paths for 2027:

• Europe absorbs it. Energy prices ease if Hormuz reopens (the IMF assumes March 2027), inflation falls back towards 3%, and protest stays in the ballot box.
• Europe does not. Energy stays high and budgets are cut to calm bond markets. The 2011 sequence repeats in a new order: bond spreads first, then the streets, then the elections. Markets would feel it in French bonds, the euro, European banks, and in North Africa and the Middle East.

7. China: two economies, and one of them rides the AI boom

China grew 4.5% in the third quarter and 4.6% in the first nine months, inside the government’s 4.5–5% target (NBS). The headline hides two economies moving in opposite directions.

The outside economy is booming, and it is the AI boom. Exports rose 35.2% in October, and the trade surplus passed $1 trillion for the second year running (CNBC).

The engine is the build-out this post follows. This year, exports of integrated circuits more than doubled, memory rose 133%, and computers and servers rose 61%. Much of that is price: memory is dear because the world is short of it.

At home, China builds AI on its own chips. Nvidia’s share there has fallen to about 8% by one estimate, and Huawei cannot make enough replacements because its chipmaker lacks the most advanced machines. Domestic chips do less work per watt, so China’s build-out needs even more power.

The inside economy is not. Retail sales of goods grew 1.1% in nine months. Consumer prices rose 0.7% in October, and 18% of young job seekers cannot find work (Caixin).

Property is in its fifth year of decline. Investment fell 20.6% and new-home sales 13.8% in nine months, even as prices in the largest cities start to flatten.

The price signal is a trap. Factory-gate prices rose 5.7% in October, the fastest since 2022, after three years of falling. It looks like recovery. It is not.

The rise comes from mining and raw materials: the war’s commodity shock reaching the factory gate, while shop prices barely move (CNBC). When costs rise faster than selling prices, margins shrink. That is a squeeze, not demand.

Beijing kept its powder dry. The 2026 budget held the deficit at about 4% of GDP and kept special-bond issuance flat. Counting everything off the budget, the deficit is about 8.1%.

The new Five-Year Plan promises a “notable” rise in household consumption, now about 40% of GDP, but gives no number.

China cushioned the oil shock with its stockpile, about 1.4 billion barrels at the end of 2025. It cut imports by roughly a third in the second quarter and drew heavily on reserves.

Our reading for 2027. This is our analysis of the numbers above, not a Kunkafa forecast, and it could be wrong.

• Growth closer to 4% than 5%. Exports cannot grow 35% again from a base that is already 35% higher, and their value rides on memory prices that turn when new factories arrive. If AI buyers slow their orders while their buildings catch up, as section 4 argues, China’s strongest engine slows with them.
• Stimulus after the slowdown, not before. Beijing has room: a 4% deficit it chose not to raise. Its habit is to spend when exports falter. That makes 2027 the year the consumption plan is tested, because exports are less likely to carry growth.
• More friction as the surplus grows. Two years above $1 trillion invite answers. The US truce runs only to January, and Europe is already weighing curbs. The yuan, near 6.7 to the dollar and rising, is the pressure valve Beijing is allowing, slowly.
• China could hold oil up even in peace. A stockpile drawn down this hard has to be refilled. If Hormuz reopens, Chinese buying to rebuild reserves could put a floor under oil just as supply returns.
• Falling prices could return when the war’s prices fade. Once oil and raw materials ease, the factory-gate rise has nothing at home behind it.

Markets feel China first through memory and chip stocks in Korea and Taiwan, through oil and metals, and through the yuan. Europe feels it in cars and machinery, where Chinese exports compete hardest.

8. Rates: the year the cuts did not come

A year ago, markets expected 2026 and 2027 to bring rate cuts. Instead the Federal Reserve raised rates in September to 3.75–4.00%.

Markets now price three more quarter-point rises by mid-2027. Goldman Sachs sees the last in March, taking the range to 4.25–4.50% (Reuters, 11 November).

October’s prices rose 0.4% in a month, and not only fuel: medical care, car insurance and airfares too (Axios).

Rates are rising, not falling: policy rates and ten-year yields, 11 November 2026

Sources: Federal Reserve H.15, ECB, Bank of Japan, Reuters, Bloomberg.

Raising rates into an energy shock has an uncomfortable history. In July 2008 the ECB raised rates as oil peaked. Within three months the world was in a financial crisis.

2027 is different, but the shape is similar. A ten-year Treasury near 5% raises the cost of every mortgage, every car loan and every refinancing, including the debt behind the AI build-out. The average new-car payment hit a record $823 in October.

Could there be a recession? The official forecasts say no. The IMF holds 2027 world growth at 3.4%, but only if Hormuz reopens by March and oil averages about $87.

If oil sits at $130 through the first quarter, the IMF’s own scenario cuts half a point off 2027 growth and lifts inflation to 5.2% (IMF).

Prediction markets give about a 28% chance of a US recession by the end of 2027. Factories are still busy, with the US ISM at 55.5 in October, but their input costs are rising at the fastest pace in years.

Our read: the risk in 2027 is less a sudden collapse than a slow grind, with rates high for longer, real incomes falling and borrowers who were priced for cuts refinancing at hikes. The recession, if it comes, is likely to arrive through credit rather than through jobs.

9. Two wars, three choke points

The war with Iran began on 28 February. An April ceasefire collapsed in July when Iran attacked commercial ships. The IEA calls it the largest supply disruption in the history of the oil market.

• Oil. Crude through Hormuz was back near pre-war volumes by late September. Refined fuels were not: about 677,000 barrels a day against 3.6 million before the war (Kpler, via CNBC).
• Still under fire. 37 ships were hit in October, the most of the war, and 10 in early November. Brent is about $92, and the IEA expects diesel to stay tight well into 2027.
• Fertiliser. The Gulf ships a quarter to a third of the world’s urea. Urea more than doubled to over $850 a tonne in April, then fell back to about $450 (WTO). About 3.9 million tonnes of exports are held up (IFPRI). Fertiliser missed this winter shows up in next summer’s harvest.
• Grain. Shipping from Ukraine’s and Russia’s Azov and Black Sea ports has largely stopped since mid-August after attacks on ships and terminals. Ukraine’s September grain exports fell 45% and wheat 55% (CNBC). Talks were due in Abu Dhabi on 10–11 November; Ukraine’s agriculture minister sees no grain truce in the coming months.

What could change US policy? The election did. Democrats won both chambers on 3 November, and analysts tie part of the result to the war and fuel prices (CNBC).

A war-powers resolution passed in June and was ignored. Another failed in the Senate by one vote in September. From January the new Congress holds the budget, and cutting war funding is its strongest tool.

The president holds a veto and has said he is weighing new strikes on Iran. Tehran reads the election as a rejection of the war, and is in no hurry.

Deal, stalemate or escalation: three paths for the Gulf war, and what each means for oil

The ranges are Kunkafa’s illustration, anchored on the IMF’s baseline and its $130 scenario, not a forecast.

The Gulf states, and the people who moved there. For a decade the Gulf sold itself as the world’s safe haven: low taxes, open visas, no wars at home. The war tested that.

In March, wealthy Asian clients asked about moving money from Dubai to Singapore and Hong Kong (Reuters), and Dubai property deals fell by a third or more. By September, private bankers told The National the trend had not reversed.

The IMF expects Saudi Arabia’s economy to shrink in 2026 and rebound strongly in 2027 if the strait reopens.

Both outcomes are open. If the war ends, the Gulf could keep its haven status, with rebuilding and energy spending lifting growth.

If it drags on, money that came from London, Moscow, Mumbai and Cairo could look for a second home, in cities with stable law and no front line nearby.

10. What people buy: cars, clothes and health

The macro risks land in household budgets.

• Cars. US sales are holding at about 15.5 million a year, but the average monthly payment on a new car reached a record $823 in October because loans cost more, not because cars do. Buyers are shifting to hybrids, now 18.6% of sales against 14.6% a year earlier, a direct answer to fuel prices.
• Clothes. H&M’s latest profit was flattered by refunds of US tariffs while its sales barely grew; Inditex grew sales but missed on profit as transport costs rose. Higher freight and fuel costs reach the price of a shirt within a season.
• Factories. Order books are healthy, but the prices factories pay are rising at the fastest pace since 2022. Either they pass them on, feeding inflation, or they absorb them, squeezing margins.
• Health. This could be the largest squeeze of all. Insurers on the ACA marketplace have proposed a median 15% rise for 2027, after enhanced subsidies expired (KFF); employers expect 8–10% (Marsh, Aon), and two-thirds of large employers plan to shift more of it to workers.

This is why “Medicare for All” is back as a campaign promise. The House bill has about 118 co-sponsors, and two-thirds of Americans favour expanding Medicare. The incoming Speaker, Hakeem Jeffries, has not endorsed it (NPR).

Cost estimates range from net savings to $47.4 trillion over ten years. With a presidential veto, it cannot become law before 2029.

For markets in 2027 it is a headline risk for insurers and hospitals. Europe’s version of the squeeze sits in its pension and health budgets, the same budgets bond markets are pricing.

11. The subscription economy stops renewing itself

For twenty years software was sold the same way: a monthly fee per person, renewed automatically, rising a little each year. Investors paid high prices for that promise because it looked safe. In 2026 it started to crack from both ends.

From the top: AI that builds the tool. On 10 November Anthropic launched its own workplace tools: Claude Sheets, Meetings, Whiteboards and Data. Sheets builds a spreadsheet from a sentence.

Anthropic itself calls it “fairly basic”, and its files export to Excel. Microsoft’s shares still fell about 2.5% that day (Bloomberg).

It was the second scare this year. In February, Anthropic’s agent tools set off a sell-off: the iShares software fund fell 24% in the first quarter, its worst since 2008.

The fear is not one product. It is the seat. Software is priced per person, and one employee with an AI agent can do the work of several. Fewer people means fewer seats to bill.

Tools are also easier to build. In Retool’s survey of 817 builders, 35% had replaced a bought tool with their own, and 78% plan to build more (Retool, via AiThority).

McKinsey reportedly found that 32% of companies dropped a software purchase because AI could build it.

From the bottom: users who say no. Households are counting. In Zuora’s survey, 47% of US consumers cancelled a subscription this year, up from 31% in 2024.

People guess they spend $86 a month on subscriptions. When they list them, it is about $219 (C+R Research). With health, car and fuel costs all rising, the monthly fee is the easiest line to cut.

The anger is sharpest at subscriptions that are easy to join and hard to leave. A court struck down the US “click to cancel” rule in July 2025, and the FTC is starting again.

The states are filling the gap. About 30 have their own renewal laws; Virginia’s took effect in July and Louisiana’s follows in January 2027. Firms that rely on customers forgetting to cancel are losing that cover, state by state.

The other side. Spending has not fallen yet. Gartner expects software spending to grow 14.6% in 2027, to about $1.68 trillion, much of it from price rises (Gartner, via Nasdaq).

In a16z’s survey, 65% of large-company technology chiefs prefer their existing vendor’s AI product, for trust and integration. Today buyers mostly renew at a lower price or refuse the increase. They rarely cancel.

So the shift in 2027 is more likely a repricing than a collapse:

• From seats to results. Vendors that charge for work done (a ticket closed, a report built) rather than per person keep their revenue as headcount falls. Vendors that cannot change lose seats at every renewal.
• From tools to data. A spreadsheet is easy to rebuild. The system that holds a company’s customers, payroll or ledger is not. The deeper the data, the safer the subscription.
• From automatic to earned. Renewal stops being a formality. A subscription has to prove its value every month, or it becomes the next line a household or a finance team cuts.

Kunkafa sells subscriptions too, and nothing in this section exempts us. Readers are right to hold every subscription, ours included, to the same two tests: is it worth keeping this month, and is leaving as easy as joining?

12. Where the opportunities could be

Every risk above moves money, and money that leaves one place arrives somewhere else. These are the places the pressure points to, not advice to buy anything.

Every risk of 2027 moves money somewhere: where the pressure creates demand

Source: Kunkafa’s analysis in this post, November 2026. Not advice to buy anything.

The order matters. The risks are near: the next quarter’s earnings, the next ceasefire, the next central-bank meeting. Most of the opportunities take a year or more to pay. That is the gap 2027 could be remembered for.

13. Conclusion: one bill, three payers

Put side by side, the risks of 2027 are one story seen from three places.

The build-out ran ahead of its power and its profit. The largest buyers own more than a quarter of their property in things that are not yet working, a trillion dollars of data-centre leases waits to start, and the labs still lose money on most of the people who use them. That is affordable while money is cheap and patient.

The wars made money dear and impatient. The squeeze on oil, fertiliser and grain pushed inflation back to about 4% in the US and Europe. Central banks answered with rate rises, the ten-year Treasury sits near 5%, and France pays the widest premium over Germany since 2012. Every borrower priced for cuts now refinances at hikes, the AI build-out included.

Households and governments pay the rest. Car payments, health premiums and fuel take more of each month. Subscriptions are the easiest line to cut, budgets the hardest to balance, and in Europe the street has already started to answer.

The summary:

• AI: prices fall about 13 times a year. Use has outrun them so far; watch compute cost as a share of revenue.
• Chips: buyers paid ahead of their buildings. Our hypothesis is that Nvidia’s growth slows more than the roughly 70% expected.
• Hosts: US permits are closing. Egypt and Morocco can win the demand if they bring power that runs at night.
• China: its growth rides the same AI boom. Our reading is closer to 4% than 5% in 2027, with stimulus arriving after exports slow.
• Money: rates are rising into an energy shock. A recession, if it comes, comes through credit.
• Wars: Hormuz and the Black Sea set the price of fuel, fertiliser and food, and the IMF’s 2027 baseline assumes Hormuz reopens by March.
• Households: health costs rise 8–15% and subscriptions are being cut. Software is being repriced, not abandoned.

What could go right. The same links work in reverse. If Hormuz reopens on time, energy falls, inflation eases and the hikes stop early.

If AI use keeps outrunning AI prices, the labs grow into their commitments and the empty buildings fill. If new hosts bring their own power, the build-out finds room.

2027 does not have to break. It just has less room for error than any year since 2008.

That is why we read every market in both directions, on every duration, every morning. When the year’s risks start to move prices, the distances show it before the headlines settle.

14. What Kunkafa’s record shows on these markets

We read the markets most exposed to these risks every day. This is how Kunkafa’s forecasts on them have turned out over their full published record: every forecast at 95% confidence or more, in both directions, by how far price travelled toward the target Kunkafa named.

Full reports: Nasdaq 100, S&P 500, Brent oil, Gold, Euro–dollar, NVIDIA, Micron, SanDisk, CoreWeave.

100% means price reached the level. It is a distance, not an investment return. The forecasts that fell short are in the same rows. Kunkafa is usually right, and the record shows exactly how often. Kunkafa is sometimes wrong, and the record shows that too. How to read forecast results explains each band.

15. Where Kunkafa stops

Kunkafa does not forecast wars, elections, central banks, recessions, economies or company results. The views on Nvidia in section 4 and China in section 7 are our analysis of public data, not Kunkafa forecasts, and nothing here is advice to buy or sell.

For the markets Kunkafa covers, Kunkafa answers one question: how far could this market move, in which direction, and how confident is Kunkafa about each outcome, across every duration at once.

16. Thirteen questions to ask in 2027

• Is AI’s compute bill falling faster than its prices? Watch compute cost as a share of revenue at every lab that reports it.
• Does Nvidia guide above $124bn for the January quarter? Below it, the slowdown hypothesis firms up.
• Are the chips switched on? Watch the share of the buyers’ property not yet in service: 30% at Alphabet and 27% at Meta, and rising.
• Who pays the real price of AI? Watch consumer plan prices, usage caps and ads: each is a price rise.
• When does memory turn? New factories, not demand, end memory shortages.
• Who grants the permit? 40% of planned US capacity sits in states promising a pause.
• Does the host bring its own power? For Egypt and Morocco, the night is the test.
• When is the last hike? Markets price March 2027; October’s prices argue for later.
• Does French debt calm down? The gap over Germany is the widest since 2012.
• Does Hormuz reopen by March? The IMF’s whole baseline depends on it.
• Does Beijing spend before or after exports slow? The answer sets China’s 2027.
• What does fertiliser cost this winter? It sets next year’s food prices.
• Is software billed per seat or per result? Watch renewals: lower prices come before cancellations.

In 2027 the arguments happen in headlines. Moves can be measured. Create a Kunkafa account and read the S&P 500, Brent oil, gold and the euro in both directions, beside their record.

Figures from Epoch AI, Anthropic, Nvidia, TrendForce, the Federal Reserve, Eurostat, the ECB, the IMF, the IEA, the FAO, KFF, Reuters, Bloomberg, the Wall Street Journal, CNBC and Kunkafa’s weekly reports of 9 November, checked on 11 November 2026.