Start with a number that should be in every conversation about AI economics and rarely is.

OpenAI, the American company behind ChatGPT, charges USD 5 for a certain quantity of AI use. A Chinese company called DeepSeek charges 14 cents for roughly the same thing. About 35 times cheaper.

Now before you start arguing with me about price and quality, I agree that a low selling price does not prove a low underlying cost. It tells you what a company charges, not what it spends.

Chinese AI companies publish no audited accounts, so nobody outside those firms knows whether 14 cents reflects genuine efficiency, thin margins, or losses absorbed by someone else.

What the price does affect, immediately, is what buyers are willing to pay anyone else.

Microsoft, Amazon, Google and Meta are collectively spending well over a trillion dollars on chips and buildings. A large part of the financial case for that spending rests on one assumption: that AI stays expensive enough, for long enough, to earn the money back. Chinese models are the most obvious external threat to that assumption, because they may put a ceiling on what the market will pay.

And the conversation about whether the AI boom makes financial sense keeps leaving that out.

Subscribe now

The bakery across the street

Imagine you own a bakery. You bought your oven expecting to sell bread for 6 years at INR 50 a loaf. That price was the whole plan. It is what pays for the oven.

Then a bakery opens across the street selling bread that customers consider good enough, for INR 1.50.

Your oven still works perfectly. Nothing about your bread has changed. But the plan you built your future on is under pressure, because the price you were counting on may no longer be available to you.

That is where the global AI industry now finds itself. And the strange part, which we will come to, is that the bakery across the street is doing it on worse equipment than yours.

This is not a fringe experiment

It would be easy to dismiss cheap Chinese AI as a curiosity that serious companies ignore. The traffic data says otherwise.

OpenRouter is a platform that tracks which AI models developers around the world actually choose when they build something. A year ago, Chinese models carried about a fifth of the traffic there. By the end of June this year, they carried nearly half. American models fell from roughly three-quarters to under a third over the same 12 months.

And many of these Chinese models are not just cheap to rent. They are free to own.

Alibaba, China’s answer to Amazon, builds a model called Qwen. Rather than keeping it locked inside its own computers the way ChatGPT is locked inside OpenAI’s, Alibaba publishes the finished model free. You can download it and run it on your own machines forever, without paying Alibaba anything. Qwen has been downloaded roughly 3 billion times worldwide in 6 months, more than any model on earth.

Giving it away is not charity and not a mistake. Alibaba still sells the computing power. Companies that build on Qwen frequently end up renting Alibaba’s machines to run it at scale, and that is where the money is.

The factories are still behind

Here the story stops being a simple tale of Chinese ascendancy.

The world’s most advanced chip manufacturing happens at TSMC in Taiwan. China’s best, a company called SMIC, is working at a level TSMC had reached back in 2018, roughly 7 or 8 years behind.

Precision is only half the problem. The other half is yield, meaning the share of what a factory attempts that comes out good enough to sell. Reported figures put SMIC’s yields on its most advanced process at 20 to 40%, and yields on Huawei’s Ascend AI chips considerably lower, as little as 5 to 20%, against 60 to 80% for the chips Nvidia has made at TSMC. Those are different figures measuring different things, and both are reported estimates rather than disclosed numbers, but the direction is consistent. China’s factories waste a great deal more of what they make.

Beijing knows this and is spending anyway. Well over USD 50 billion of state money has gone into expanding these factories. This is not yet an industry that pays for itself. It exists because the Chinese government has decided the country cannot afford to be without it.

At the aggregate level the gap is large. The Council on Foreign Relations estimates that the United States and its allies can produce roughly 35 to 38 times more advanced AI chip manufacturing capacity than China can, once adjusted for quality and for how many Chinese chips actually work. That is a measure of capacity to manufacture leading-edge chips, not a count of how many AI chips each side possesses.

The domestic chips are doing real work

Chinese engineers are getting meaningful results from that inferior equipment.

Huawei’s Ascend chips were used for part of the training of DeepSeek’s V4-Flash model. Separately, a Huawei-led team, working with 3 Chinese research institutes, completed the post-training stage of the larger V4-Pro model on a cluster of at least 1,000 Ascend chips.

That distinction is important. Building an AI model happens in stages. The initial training is the hardest and most expensive part. Post-training is the refinement that follows, and running the finished model afterwards is easier still. Chinese domestic chips were widely assumed to be much less capable of handling the hardest stages of frontier-model development. They are now doing more than that.

Two things are true at once.

1. China’s chip factories are years behind and waste far more of what they make.

2. Chinese engineers are extracting serious work from that equipment regardless.

Two scoreboards, moving in opposite directions

On model quality, China has nearly caught up. Stanford University’s AI Index tracks the gap between the best American and Chinese models on standard tests. Three years ago it sat somewhere between 17 and 32 points. It is now under 3.

On manufacturing capacity, the gap has not moved, and some estimates have it widening over the next 2 years.

• If China’s models were far worse, cheap Chinese AI would only ever compete for low-value work, and premium pricing would hold everywhere else.

• If China’s hardware constraint were binding today, cheap Chinese AI could not scale to meet serious demand.

Neither is currently the case. The models are close enough to compete, and the hardware, while worse, has not yet stopped them. That is what allows a price set in China to influence what buyers everywhere expect to pay.

Three explanations, and no way to choose between them

Nobody outside these companies can say why Chinese AI is priced where it is. There are 3 candidate explanations.

1. Engineering efficiency. Chinese labs may be extracting far more from each chip through better software, the way a skilled cook gets more from a cheap stove than a poor one gets from an expensive kitchen.

2. Subsidy or absorbed losses. With no audited accounts published, it is not possible to tell from outside whether these prices cover costs, or whether the state or the parent company is carrying the difference.

3. The constraint has not bitten yet. The hardware gap may not have mattered so far because nobody has needed to push these chips to their real limit. That can change.

We do not know which of these, or which combination, explains the observed prices. The answer will determine whether the pricing pressure is durable or temporary.

The obvious escape hatch

There is a serious counter-argument to all of this. Falling prices can create so much additional demand that total revenue rises anyway.

Cheaper AI means more uses become affordable, volume expands, and the infrastructure gets used harder than anyone planned. This is Jevons’ paradox, and I have written about it at length in an earlier piece in this series, including the arithmetic showing revenue rising sharply even as unit prices collapsed.

The question is whether demand grows faster than prices fall. And whether it grows fast enough for each individual chip, not just for the industry as a whole. A market can double in size while the machines inside it still fail to earn back what they cost.

What this actually means

Come back to the trillion dollars. I am not predicting that the American AI buildout collapses. The machines work, the demand is real, and the escape hatch above is genuine.

BUT, the return on all that capital looks considerably less comfortable if the price of AI keeps falling, and China is the most visible reason to think it might. That pressure comes from outside the system entirely. A competitor selling at a fraction of your price does not care which depreciation schedule you chose, or who guaranteed your loans. It only has to keep selling.

For years the AI race has been reported as a scoreboard: who has the smartest model, who is winning.

China adds a spicy plot twist. A country whose factories are churning out older-generation hardware with much lower yields should not, on the face of it, be able to support software good enough to influence global prices. But this is precisely what is happening. And much of the West’s investment case assumes prices remain high enough to justify the scale of the investment.

The oven those American companies bought still works perfectly. That was never in doubt. What is in doubt is the price it can earn back.

Leave a comment

Share

Subscribe now

Sources

Pricing: OpenAI and DeepSeek published price lists, as at the dates cited. These change frequently and should be checked against the providers’ own pages before reuse.

Traffic share: CNBC, reporting OpenRouter data for the final week of June 2026, showing Chinese models at 48% of tracked traffic against 32% for US models, up from roughly 20% and down from roughly 74% respectively a year earlier.

Qwen downloads: Alibaba’s own stated figure of roughly 3 billion global downloads over 6 months, as reported by Bloomberg. Hugging Face’s 2026 platform figures show a narrower like-for-like count nearer 2 billion on that platform specifically. These measure different things and should not be merged.

DeepSeek V4 and Huawei chips: Reuters and Tom’s Hardware, April and June 2026, reporting Huawei Ascend chips used in part of the training of DeepSeek V4-Flash, and a Huawei-led team completing post-training of DeepSeek V4-Pro on a cluster of at least 1,000 Ascend 910C chips in partnership with the Shenzhen Loop Area Institute, Harbin Institute of Technology’s Shenzhen campus, and the Shenzhen Research Institute of Big Data.

Manufacturing and yields: reporting compiling Council on Foreign Relations analysis and industry sources on SMIC’s continued 7nm process versus TSMC’s more advanced nodes; reported SMIC 7nm yields of 20 to 40%; Huawei Ascend yields reported as low as 5 to 20% against 60 to 80% for Nvidia’s TSMC-manufactured chips. These are reported estimates from industry sources, not figures disclosed by the companies.

Compute and manufacturing capacity gap: Council on Foreign Relations, “China’s AI Chip Deficit,” on the estimated 5-times current performance gap between the best US and Chinese AI chips, widening to 17 times by 2027, and the separate estimate of US and allied manufacturing capacity for advanced AI chip dies at 35 to 38 times China’s, quality-adjusted.

Model capability gap: Stanford HAI’s AI Index, as cited in reporting, on the narrowing gap between top US and Chinese models on standard benchmarks, from an estimated 17.5 to 31.6 points in 2023 to under 3 points by early 2026.

State funding: reporting on Chinese national semiconductor fund allocations to SMIC’s expansion exceeding USD 50 billion.

For the last few weeks, my phone would not stop talking to me about Nvidia.

Subscribe now

I had recently developed an unhealthy interest in the money sloshing around AI companies, and every news app I own took that as permission to feed me Nvidia headlines around the clock. Record earnings. A USD 500 billion financing deal. A guarantee here, an investment there. The same name, over and over, until the obvious question surfaced.

Why did this one company matter so much?

So I started digging. And the deeper I went, the more a second question crowded out the first, one I could not un-see. What on earth Nvidia had done to become this foundational to an entire industry.

I am going to answer both, but in reverse order. Because to understand how Nvidia became the juggernaut it is today, you first have to understand what it did to get here.

First, picture what Nvidia has become

Picture one company that owns the flour mill, lends the local bakeries money to buy ovens, takes a stake in several of those bakeries, and then guarantees their rent to the landlord. If that company stumbled, it wouldn’t just lose money. It would take every bakery in town down with it.

That is roughly the position Nvidia has built for itself in AI. It happened in stages, each one adding to the last, until the company that makes the chips became the foundation the entire industry stands on. To see why that is either brilliant or alarming, depending on who you ask, you need to know how Nvidia made itself impossible to avoid in the first place.

The 20-year bet almost nobody understood at the time

Nvidia does not actually make its own chips. It designs them.

The physical manufacturing happens in Taiwan, at a company called TSMC (Taiwan Semiconductor Manufacturing Company), and we’ll come back to why that matters. Nvidia’s real head start came from something you cannot touch.

In 2006, Nvidia released software called CUDA (Compute Unified Device Architecture).

Think of a graphics chip, a GPU (Graphics Processing Unit), as an oven originally built to do one job fast: bake large loaves of bread. A GPU chip was designed to draw video game images by handling thousands of tiny calculations at the same time. CUDA was the thing that let anyone use that same oven to cook something other than large loaves of bread. Think croissants, cupcakes, et all. CUDA made smaller individual units of output produced much faster.

With CUDA, a chip built for games could crunch any problem that could be broken into thousands of parallel pieces. Nvidia gave CUDA away free, taught it in universities, and built a generation of programmers who knew no other way of working.

The man who built CUDA, Ian Buck, said this year that it cost the company billions and made no money for 10 years. Nvidia never dropped it.

Then luck met preparation. It turned out that the maths behind modern AI, the kind that trains a model by doing enormous numbers of small calculations at once, is almost exactly the kind of maths Nvidia’s chips had been built to do for video games. When AI took off after 2012, Nvidia was not scrambling to catch up.

A generation of AI researchers had already built their work around CUDA.

A competitor can build a fast chip, but it cannot easily rebuild 20 years of software, libraries and trained engineers wrapped around Nvidia’s. That is the moat, and it is why, when the AI boom arrived, there was really only one place to buy the equipment. Nvidia.

Nvidia has one genuine dependency. TSMC. The hardest part of making these chips is not the chip itself, it is the packaging: stacking the processor next to its memory with a precision no other factory can match at scale. Nvidia designs the miracle. TSMC is the only company currently producing it at anything like the scale Nvidia needs.

Now, what it started doing with that grip

For most of this story, Nvidia’s job was simple. It sold the GPUs (ovens, in the example). People bought the GPUs. A normal business, operating at a scale that is genuinely hard to picture.

On 26 August, Nvidia reported USD 96.2 billion of revenue for a single 3-month period, more than double the same quarter a year earlier. It expects USD 108 billion in the next 3 months. About 92% of that comes from selling the chips that power AI data centres.

That alone would be a remarkable story. It isn’t the whole one, because Nvidia stopped simply selling ovens and started financing the bakeries too.

Nvidia starts lending money

A data centre full of Nvidia’s best chips costs billions. Most of the companies that want to build one, including the most famous AI names on earth, do not have that money sitting in the bank. So Nvidia started investing directly in its own customers, handing them cash to spend on Nvidia’s own chips.

On 26 August its finance chief, Colette Kress, told investors the company has now put nearly USD 50 billion into what it calls “frontier AI labs,” the small handful of firms, OpenAI and Anthropic among them, that build the most advanced models in the world.

Picture the flour mill lending the bakery the money to buy the oven, on the condition that the oven is bought from the mill. The bakery gets its equipment. The mill gets a guaranteed sale, and often a slice of the bakery’s future profits on top. One company - controlling the entire chain. Because it controls the equipment and has the money.

Then it started guaranteeing their loans

Investing cash was the easy part. The bigger move came next, and Jensen Huang, Nvidia’s chief executive, explained exactly why it was needed, in his own words, on that same call.

The frontier labs, he said, have “extraordinary demand for training and inference compute,” but they are “growing faster than what their balance sheets and credit profiles can support.”

Ergo: business is booming, but the banks won’t lend them what they need, because these companies haven’t been profitable long enough, or reliably enough, to earn the kind of credit rating that unlocks decade-long loans.

So Nvidia started signing for them.

OpenAI agreed to a giant AI campus on a former nuclear site in Pike County. OpenAI cannot get an investment-grade credit rating on its own, because it has not turned a profit, and lenders were wary of financing something this size for a company with that history. So Nvidia agreed to stand behind the lease and power payments. If OpenAI cannot pay, Nvidia will.

Watch the number attached to that promise, because it moved 3 times in a single month, and every move is on the public record. In July, reporting put it near USD 250 billion. By mid-August, Nvidia had cut it to under USD 120 billion, after its own investors balked at the risk. When the deal was signed on 17 August, the guarantee landed at USD 105 billion, filed with America’s market regulator as a capped payment obligation.

Three figures, each smaller than the last, the second-biggest company on earth being talked down in public over 10 days.

Nvidia didn’t stop there. It put USD 1.5 billion of its own money into SB Energy, the company building and owning the site. So, on this one project Nvidia is the:

1. Chip supplier

2. Investor in the tenant

3. Investor in the landlord

4. Guarantor of the rent.

One company, 4 positions, same deal.

Fortune, reporting the reduced guarantee, named the worry directly: it reflects “the circular cycle of money in the AI ecosystem.”

Then it pulled in everyone else’s money too

One company backing its own customers is risky enough. The next step spread that risk across some of the largest pools of money in the world.

In August, Nvidia announced financing partnerships with 6 of Wall Street’s biggest firms: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. The plan is to raise more than USD 500 billion from outside investors to fund AI data centres. And Huang used the moment to declare something new. Nvidia’s chips, he said, are now an “investable asset class.“

He wants pension funds and insurers, the cautious giants that normally lend against office towers and toll roads, to start lending against stacks of graphics cards instead.

Kress saw the obvious objection coming and got ahead of it. “We recognise the scale of this support,” she told analysts, “and we know some will call this circular financing. We see it differently.” Huang went further when pressed. He called the labs “once-in-a-generation companies,” and added, “the only regret that I have is that I didn’t invest more and sooner.” He may be right.

Vendor financing, a company helping customers afford its own products, is neither new nor illegal. Cisco did it in the 1990s. General Electric built a whole finance arm on it. Kress also has a specific, fair defence of the second, still-unnamed lab Nvidia is backing: its credit support only tops up money that lab had already raised on its own, rather than replacing it.

Where the two halves of the story stop fitting together

The tension sits inside Huang’s own words, spoken on the same call, minutes apart.

On one side, he said AI has “reached its inflection point,” that it is “doing useful work,” that its tokens are “productive and profitable,” and that “compute is revenue.” A business, in other words, that works. On the other, he said the very companies driving that boom cannot get a bank to lend to them, because their finances don’t yet look safe enough.

A business that is genuinely turning profitable does not usually need someone else to co-sign its loans.

Banks lend against real, growing profit every day. If the profits Huang describes were convincing lenders on their own, the guarantees would not be necessary. That they are necessary is itself the tell.

The people whose entire job is pricing risk are not yet as sure about these companies as the earnings call wants you to be.

Epilogue:

Nvidia now touches nearly every stage of the AI money cycle, in a way almost no company in history has gripped a single industry.

• It makes the chips.

• It lends the money to buy the chips.

• It owns pieces of the companies buying them.

• It guarantees their largest loans.

• It is now recruiting the most cautious investors on the planet to lend against the very equipment it makes.

None of this is hidden. Nvidia lays it out in its own filings and on its own earnings calls. Huang has said, on the record, that his biggest customers cannot finance this buildout without him.

Earlier this year, Nvidia’s own filings put the total of every building-lease guarantee it had ever given at USD 3.5 billion. The Ohio guarantee alone is USD 105 billion.

30 times that entire prior history, in one deal.

The company went from occasionally backstopping a partner’s lease to underwriting a single project worth more than a hundred billion dollars, inside 12 months.

Whether that turns out to be patient, once-in-a-generation conviction, or the industry’s biggest company holding up demand for its own product, no one outside these boardrooms can say yet.

What can be said is this. To judge whether the AI boom rests on solid ground, it is no longer enough to ask whether people want to use AI. You now have to ask whether one company’s balance sheet can keep underwriting almost everyone else’s. That is a very different question.

Leave a comment

Share

Subscribe now

Sources:

Q2 FY2027 results: NVIDIA’s official press release and Form 8-K, 26 August 2026, for revenue of USD 96.2 billion (up 106% year on year), GAAP gross margin of 75.0%, GAAP and non-GAAP earnings per diluted share of USD 2.46 and USD 2.22, Data Center revenue of approximately USD 89 billion (about 92% of total revenue), and the Q3 FY27 revenue guide of USD 108 billion.

CUDA and Nvidia’s origins: Ian Buck’s remarks at GTC 2026 on CUDA costing the company billions and making no money for a decade, as reported by Computer Weekly (March 2026); multiple technology-industry histories of CUDA’s 2006 release, its free distribution to universities, the developer lock-in effect, and the growth of the CUDA developer base from 1.8 million in 2020 to 4.5 million; the 2012 AlexNet breakthrough as the point at which GPU-based deep learning took off. The parallel-processing point (that the matrix mathematics of deep learning matches the parallel design of a graphics chip) is a widely documented technical explanation of Nvidia’s AI fit.

TSMC dependency: industry reporting on Nvidia’s fabless model, its exclusive reliance on TSMC for advanced Chip-on-Wafer-on-Substrate (CoWoS) packaging of its H100 and Blackwell chips, and TSMC’s unmatched yield rates at that scale.

Huang and Kress quotes: NVIDIA’s Q2 FY2027 earnings call transcript, 26 August 2026, for Huang’s statements on frontier AI labs outgrowing their balance sheets and credit profiles, the “inflection point” and “compute is revenue” framing, the “once-in-a-generation companies” and investment-regret comments, and Kress’s confirmation of nearly USD 50 billion invested in frontier AI labs, the 6-firm financing partnerships targeting over USD 500 billion in third-party capital, and her direct response to the circular-financing characterisation.

The Ohio guarantee: reporting on the evolving Nvidia-OpenAI Ohio financing structure through July and August 2026 (TechPowerUp, Yahoo Finance, Fortune), tracking the guarantee from a discussed roughly USD 250 billion, to a reduction to under USD 120 billion by 14 August, to the signed figure; OpenAI, Nvidia and SB Energy’s joint announcement and accompanying SEC filing, 17 August 2026, confirming the final capped “aggregate payment obligation” of USD 105 billion, the 20-year, 8-gigawatt lease at the Pike County, Ohio site, and Nvidia’s USD 1.5 billion direct investment in SB Energy. Fortune’s characterisation of the reduction as reflecting “the circular cycle of money in the AI ecosystem,” 18 August 2026.

Nvidia’s prior guarantee exposure: NVIDIA Form 10-Q for the quarter ended 26 April 2026, Note 8, disclosing maximum gross exposure of USD 3.5 billion across all facility lease guarantees given to that point, used as the baseline against which the Ohio guarantee’s scale is compared.

You have been investing since your first salary. The SIP (systematic investment plan, a fixed monthly contribution into a mutual fund) is running. The EPF is growing. There is an ELSS fund in there somewhere, started the year someone explained tax deductions to you over lunch.

All of it is good. All of it was designed for a smaller version of your financial life. At your new income you can afford to build in layers, where each one protects the layer above it.

Subscribe now

Layer 1: protection

This comes first, because a single uninsured event can undo 10 years of disciplined investing in an afternoon.

• Your emergency fund should cover 6 months of your current essential expenses. Recalculate it against what you spend now, not what you spent before the promotion.

• Your health cover needs a fresh look. A serious hospitalisation in a metro can run well past what most employer policies carry, and employer cover disappears on your last working day, exactly when you may be between jobs and least able to buy new cover. A personal base policy plus a super top-up (a policy that pays only above a threshold, and therefore costs a fraction of equivalent base cover) closes that gap cheaply.

• Term insurance should sit at 10 to 15 times annual income, recalculated at every material income change, and it matters most if anyone depends on you financially.

• Income protection, which pays a monthly income if illness or disability stops you working, is the most underused product available to salaried professionals in India. Your ability to earn is the asset funding everything else on this list.

Layer 2: growth

• Increase your index fund SIP by at least 50% of the net increment, and do it in the first week, before the money learns a new route.

• If you are on the old regime with unused Section 123 room after EPF, fill it with ELSS. On either regime, employer NPS under Section 124(2) is worth pursuing.

• Add consistency rather than complexity. A boring portfolio that runs untouched for 20 years reliably beats an exciting one restructured every 18 months, because the restructuring is where the returns leak out.

Layer 3: Freedom

• International index funds for geographic diversification.

• Goal-linked SIPs for specific milestones with dates attached.

• Liquid funds instead of a savings account for money that is waiting for a decision. This layer is optional at this stage. Build it once the first two are unshakeable.

What to stop doing this week

• Holding endowment policies and ULIPs and calling them investments, when the charges rarely justify the returns and the insurance component is thin.

• Letting increments accumulate in a savings account awaiting a decision that never gets made.

• Buying last year’s top-performing fund.

• Holding large fixed deposits beyond your emergency fund.

• Taking product recommendations from relatives at weddings, who mean well and sold you an endowment policy in 2019.

Protection, then growth, then freedom. In that order, every time.

What to do right now:

1. Recalculate your emergency fund: current monthly essential expenses multiplied by 6. Close any gap within 3 to 6 months through a standing transfer.

2. Increase your SIP by at least 50% of the net increment, in the first week.

3. Check your Section 123 utilisation if you are on the old regime. Subtract EPF from 1,50,000 and fill only the remainder.

4. Ask HR about employer NPS under Section 124(2). It works on both regimes.

5. Review your health cover, and hold a personal policy rather than depending on your employer’s.

6. Recalculate term cover against your current income and liabilities.

7. Review every product you own. If you cannot explain in one sentence what it does and what it costs you annually, that is the one to investigate first.

8. Update nominee details on every policy, folio and account.

Leave a comment

Share

Subscribe now

This is not an exaggeration or a misremembered statistic. In the financial year 1973-74, India’s income tax had 11 slabs, climbing to a statutory top rate of 85% on income above ₹2 lakh. On top of that sat a surcharge of 15%, levied on the tax itself. Stack the two together and the effective top marginal rate reached 97.75%. The state took almost all of it, and left the earner the change.

It is easy, from here, to laugh at this as obvious madness. It is more useful to understand why a serious country, full of serious people, did it, and what happened next, because what happened next is the most important lesson in the entire four-part story.

Subscribe now

Why a free nation taxed itself to the edge

India after independence chose a particular path. It was a planned, socialist-leaning economy, and high marginal tax rates on the rich were not an accident of that choice. They were the philosophy of it made real. The intellectual scaffolding had been built in the 1950s, when a Taxation Enquiry Commission recommended a top rate as high as 85% on the largest incomes, describing it as about as far as one could reasonably go. By the early 1970s, with surcharges piled on, the system had gone considerably further than that.

The goal was equality. Wealth concentrated at the top would be taxed down toward the rest. It is a coherent moral idea, and it failed, comprehensively, for a reason that the Arthashastra could have predicted 2,300 years earlier.

At a 97.75% rate, the state was no longer collecting revenue. It was manufacturing avoidance. When declaring your income honestly leaves you with barely any remnant, the rational response is to stop declaring it. So people did, on a massive scale.

India in this era ran a parallel “black” economy that some estimates put as larger than the formal, taxed one. The confiscatory rate did not redistribute wealth. It drove wealth underground, corroded tax morality, and turned evasion from a crime into a norm. The state reached so greedily for the nectar that it killed the flower, exactly as Kautilya warned.

The correction, when it came, proved the point. A committee recommended cutting the top rate from 97.75% to 75%, and over the following two decades successive finance ministers kept bringing it lower. By the “dream budget” of 1997, the peak rate of personal income tax was 30%.

And as the rates fell, collections rose. Lower rates, better compliance, a wider base, more revenue. The country spent 25 years learning, painfully, the lesson Kautilya had spoken about.

The taxes that came, and went

The income tax was only the headline. Independent India also experimented with an entire menu of other taxes, and the fate of two of them is worth a close look.

Wealth tax, introduced in 1957::

• Where income tax falls on what you earn, wealth tax fell on what you had: the net value of your assets above a threshold, taxed every single year. It expressed the same equalising instinct as the high income tax rates. And it ran into the same wall, in a different form. Valuing a person’s entire net worth annually is enormously laborious, the kind of thing that requires jewellery to be appraised and property to be assessed, year after year.

• By the end, wealth tax was collecting around ₹1,000 crore a year, under 0.1% of total tax revenue, for all that trouble. From 1957 through the 1990s, no reliable year-wise administrative-cost-as-% series exists in the public domain; for the period for which data does exist (2001-02 onwards), the ratio was roughly 54% in 2001-02 and fell to about 9% by 2013-14.

• Wealth Tax was abolished in 2015 and replaced with an additional 2% surcharge [a charge levied on the tax itself] on incomes above ₹1 crore, which the government expected to raise about ₹9,000 crore. Roughly 9 times the money, from a levy that required nobody to value a necklace.

Estate duty, a tax on inherited wealth, introduced in 1953.

• The principle was old and respectable: tax the transfer of fortunes between generations, both to raise revenue and to soften the inheritance of advantage. It was repealed in 1985, and the finance minister who killed it gave reasons that should sound familiar by now.

• It had not meaningfully reduced inequality, it raised very little, around ₹20 crore a year by the end, and it cost a great deal to run. The thread connecting these abolitions is practical, not ideological: the unglamorous discipline at the heart of all good tax policy.

A tax that costs more to collect than it brings in, or that drives the behaviour it taxes into hiding, is a bad tax, however noble its intent. India learned this the long way, by trying.

The thicket, and the great simplification

If you have ever looked at an Indian tax calculation and felt the ground shift under you, you are not imagining it, and you are in good historical company.

Independent India developed a genuine thicket of stacked levies. There was the base tax, and then the surcharge, a tax levied on the tax. Then came cesses, additional charges earmarked, at least in theory, for specific purposes, an education cess, later a health and education cess, each one a small charge levied on the tax-plus-surcharge.

The structure compounds: a slice of a slice of a slice. It is the kind of thing that is easy to mock, and more useful to understand, because each layer was added for a reason that made sense to someone at the time, and the sum of all those reasonable additions is the complexity every Indian taxpayer now navigates.

The indirect tax side was, for decades, even more tangled. Every state levied its own taxes on goods, with its own rates and rules, on top of central levies, so that a product moving across India was taxed and re-taxed at every border, and a tax was often charged on a price that already included an earlier tax. It was the cascade problem that Maurice Lauré had designed VAT to solve in France in 1954, and India lived with its worst version for a very long time.

The answer arrived in 2017, with the Goods and Services Tax. GST folded a tangle of central and state levies into one system built on the value-added principle. Tax is charged at every stage, but each business claims credit for the tax already paid down the chain. The cascade breaks. It was the largest tax reform in independent India’s history, and the destination of this entire series.

The value-added idea that a French official sketched in 1954, itself descended from transaction taxes that Rome and Uruk would have recognised, finally unified the indirect taxation of a subcontinent.

Where the whole story leaves you

Stand back from all four parts and the long arc is clear.

For 5,000 years the state taxed what it could see, because it could not see what you earned. Grain, salt, trade, land, the transaction at the point of sale. Then, in the last 200 years, the state learned to see income, and the income tax moved from a desperate wartime expedient to the centre of public finance.

India compressed an extraordinary amount of this history into a single human lifetime: from a 97.75% rate that taught the country what overreach costs, through the slow abandonment of taxes that cost more than they raised, to a GST that returns, at the end, to the oldest idea in the book. Tax the transaction.

The reason any of this matters to you, specifically, is that you are living inside the latest chapter, and it is still being written. Every Budget speech you half-listen to is another adjustment in the 5,000-year-old negotiation between a state that needs money and a people who would rather keep theirs. The rates change. The names change.

The fundamental tension, between what the state can take and what the people will bear, has not changed since the first scribe in Uruk pressed a number into wet clay and recorded what somebody owed.

This is the final part of the tax series.

Series Guide - History of Taxes:

• Part 1: What the First Writing Was For

• Part 2: The Tax That War Built

• Part 3: Kautilya to Gandhi

• ✓ Part 4: The Most Expensive 100 Rupees in History

Sources:

• Business Standard, “The 70-year journey of income tax in India,” and Business Today, “From 97.75% in 1973 to 42.74%,” on the 1973-74 slab structure (11 slabs, 85% top statutory rate, 15% surcharge on the tax for incomes above ₹15,000, 97.75% effective top marginal rate); on the Wanchoo Committee recommendation to cut the top rate from 97.75% to 75%; on the 1997 “dream budget” 30% peak rate

• Standard accounts of the Taxation Enquiry Commission (1953-54) and the 85% recommendation

• Wealth-tax Act, 1957, and Finance Act 2015, on the introduction (1957) and abolition (2015) of wealth tax; 2013-14 yield of approximately ₹1,008 crore, under 0.1% of total tax revenue

• Estate Duty Act 1953 and its 1985 repeal: approximately ₹20 crore annual yield against high administration cost; reasons stated by the government at abolition

• Standard accounts of surcharge and cess structure in Indian direct taxation

• Goods and Services Tax (2017): the unification of central and state indirect taxes on a value-added basis

• Wealth-tax Act, 1957, and Union Budget speech 2015-16, on the introduction (1957) and abolition (2015) of wealth tax; 2013-14 yield of approximately ₹1,008 crore, under 0.1% of total tax revenue; the replacement 2% surcharge on incomes above ₹1 crore and its targeted collection of approximately ₹9,000 crore

Subscribe now

Share

Leave a comment

I want to dismantle something before we go any further. The idea that rest is something you earn. That recovery is a reward for sufficient effort, a treat you allow yourself after you’ve done enough, logged enough hours, crossed enough things off the list. That guilt about stopping is a sign of conscientiousness rather than a sign that something has gone wrong in how you understand effort and its relationship to rest.

Here is what I actually know about recovery after a lifetime of doing difficult things in difficult circumstances. Rest without guilt is available the moment the effort is honest.

That’s it. That’s the whole secret. And it is both simpler and more demanding than anything the wellness industry will ever sell you.

Guilt about rest has a very specific source.

It lives in the gap between the effort you made and the effort you know you were capable of. It grows in the space where you cut a corner, renegotiated a decision you had already made, held something back that you had available to give. When the work was incomplete or dishonest, the rest feels stolen. Part of you knows the tab is still open. And open tabs don’t let you rest, not really, not in any way that actually recovers anything.

But when you have genuinely done your best. When the effort was complete and honest and you left nothing useful on the table. The rest isn’t an indulgence or a reward or something that requires justification. It’s just the next thing in the sequence. Identify. Complete. Next. Rest. Repeat.

The rest was always in the sequence. It arrives without guilt when it follows genuine effort because there is nothing for the guilt to grip.

The problem was never the rest. The problem was always the incomplete effort that preceded it.

I should tell you what rest actually looks like from the inside. Because it is probably not what you’re picturing.

It doesn’t always mean sleep. It doesn’t require a yoga mat or a meditation app or a carefully curated wind-down routine with specific lighting and the right kind of herbal tea. It doesn’t need a name or a protocol or a wellness framework built around it.

Rest means closing the open tab so you can move to the next thing. And the next thing can be anything. Anything at all, provided you bring the same quality of presence to it that you brought to the work.

For me, it is sometimes feeding the crows.

I want to pause on that for a moment because it deserves more than a passing mention. Crows are not incidental birds. They are intelligent, observant, and each one is capable of teaching Loki himself a thing or two about mischief. They are, in other words, entirely appropriate company for someone who has spent the day doing difficult things with full attention and is now stepping outside to do a small, complete, intentional act of feeding before moving on to whatever comes next. The crows get the same quality of presence as the work. Not a distracted scattering of seeds while mentally composing tomorrow’s to-do list. Full attention. Complete intention. The tab that was open is now closed and this, right now, is the only tab running.

Sometimes rest looks like laughing at unhinged Pinterest boards with my daughter and laughing till the tears flow (oh my eyes, MY EYES). Which I do not recommend explaining to anyone who asks what you’re doing because “recovering intentionally via algorithmically curated chaos” sounds considerably less dignified than it feels.

But that’s the point. When you do everything with intention, including the laughter, even these become spaces to catch your breath. The comedy of the completely unhinged is as restorative as sleep when you arrive at it fully, without one eye still on the work you left behind.

Rest is not the absence of activity. It is the presence of intention in whatever you’re doing.

There is a reason athletes understand this better than most.

The training creates the strain. The recovery is what actually builds the strength. You do not get stronger during the workout. You get stronger after it, when the body is given space to rebuild what the effort broke down. Skip the recovery and you don’t get more fit. You get injured.

The mind is not built differently. It just doesn’t come with the same obvious warning signals. A muscle tear announces itself immediately and loudly. Mental depletion sneaks in on padded feet, more gradual, easier to ignore until the day you sit down to work and discover that the spotlight won’t switch on properly and the staccato rhythm has gone flat and everything that used to feel clear feels slightly foggy and resistant.

That’s not a character flaw. That’s a body telling you the recovery was skipped too many times.

The solution is not to push through. The solution is to close the tab properly, go feed the crows, laugh at something completely ridiculous, sleep when sleep is available, and come back tomorrow with a mind that has actually been given space to do what minds do when you stop demanding things of them.

They sort. They connect. They make sense of things that felt messy when you were in the middle of them. They find the solution you were forcing that refused to come, and they hand it to you casually while you’re standing outside watching an intelligent, Loki-adjacent bird regard you with knowing amusement.

I want to return to the guilt, because it deserves a direct answer.

If you are someone who finds rest genuinely difficult, who feels the pull of the unfinished even when you’re supposed to be recovering, who cannot quite switch off because the list is always there in the background doing its quiet damage, I want to ask you one honest question.

Was the effort complete?

Not perfect. Complete. Did you do your best, with what you had, from where you were, without renegotiating the decision halfway through? If yes, the guilt has no basis. It is a habit, not a signal. And habits can be retrained.

If no, the guilt is information. It is telling you something useful about the gap between the effort you made and the effort you were capable of. Listen to it, close the tab properly tomorrow, and then rest without apology.

The crows will be there either way. They are patient, intelligent birds. They can wait.

Share

Subscribe now

My son was stuck on his maths homework a few weeks ago. He had been at it long enough to be properly fed up, and he looked up and asked me who invented numbers. I had nothing.

Subscribe now

I had a vague sense that zero came from somewhere around Persia, possibly, and after that the well was dry. That is a poor answer from anybody. It is an embarrassing one from me, given that I have spent 2 decades in finance and write about numbers most weeks of the year. So, I went looking. Several nights of ruined sleep later, this is what I found.

Two things surprised me. My hazy memory about Persia was wrong in an interesting way. And zero, which is the part everybody half-remembers, is not the invention that mattered most.

A note before we start. This story comes from archaeology and from scholars arguing about archaeology. There are no filings here, no regulatory orders, nothing I can put in front of you and say look, page 14. Where the experts disagree, I have said so and put the names in the footnotes. Where the dating is uncertain, I have given you the range instead of picking a number and sounding confident.

What we start with:

We are born able to count to about 3. Show anyone 1, 2 or 3 objects and they know how many there are instantly, without counting. Psychologists call this subitizing, the ability to recognise a small quantity without counting it [FN1].

Go past 4 and it stops working. Babies can do it. So can crows and bees.

Everything above 4 had to be built. You need a fixed run of words, said one per object, where the last word tells you the size of the whole pile. That final step is harder than it looks. Plenty of small children can chant numbers up to 20 and still not be able to tell you how many things they just counted.

The oldest objects that might be counting are notched bones. The Lebombo bone, found on the border of South Africa and Eswatini, carries at least 29 cuts and is somewhere between 42,000 and 44,000 years old. The Ishango bone from the Congo is around 20,000 years old and carries 3 columns of grouped notches, which different scholars have read as arithmetic, as a calendar, and as modern mathematics being projected onto marks whose purpose remains unknown [FN2].

A notch does not come labelled with its purpose. Someone made deliberate, repeated, grouped marks. What they were for is still open.

Why we count in tens:

The base of a number system is the number it groups quantities around before starting a new place. We use base 10, which is why 9 rolls over into 10.

Almost every system that survived is built on 10 or 20, and the reason is at the end of your arms. Base 10 follows the fingers. Base 20 follows fingers and toes together, and it is still visible in French, where 80 is quatre-vingts, 4 twenties.

Base 60 is the odd one out, and it is the one the Mesopotamians chose. Nobody knows why. One explanation is that 2 older measurement systems merged. Another is that 60 is unusually useful because it divides cleanly into so many numbers [FN3]. A third involves counting finger joints. All 3 are plausible. None is proven.

You still use base 60. It is on your wrist, and in every angle you have ever measured.

A + Counting = Accounting:

The earliest writing we have comes from Uruk, in what is now Iraq, from around 3300 BCE. It is almost entirely accounting. Who owed what to whom.

Those scribes did not use 1 number system. They used 15, at the same time, and picked between them according to what they were counting [FN4]. Grain ran on one. Fields ran on another. Sheep, people and jars ran on a third. Rations of bread and fish ran on a fourth.

As if this wasn’t confusing enough, the same mark meant different amounts in different systems. There is a sign that means 1 unit on a tablet about sheep or jars, 1 unit again on a tablet about bread rations, 5 sila (a grain capacity measure) on a tablet about grain, and 1 iku (a unit of field area) on a tablet about land. A second sign, normally worth 10, is worth 18 when it sits beside the sign for surface area.

The mark did not carry a number that you then applied to a commodity. Its value came from which system it was sitting in. This is genuinely hard, and not only for us. Specialists read the grain system as decimal for decades, until Jöran Friberg showed in the late 1970s that it was nothing of the kind.

Compare that to now. When you write 5 kg and 5 km, both contain the same 5. The unit of measurement is added afterwards, and you can take it off. Strip the units and you can add the numbers.

In Uruk there was nothing to strip. Within any one system you could count and compare perfectly well, so a scribe knew that 6 sheep beat 5 sheep. Across systems there was no way to compare anything, because there was no general number standing behind the marks that both systems could be converted into.

You could not have run a trial balance in Uruk. A trial balance assumes every figure on the page is the same kind of figure and can be added into one total. In Uruk they were not. The number was stuck to the thing.

The Big Split:

Over the next thousand years, things came unstuck.

By around 2100 to 2000 BCE, scribes had started doing something new. They would take a quantity out of its measurement system, turn it into a plain number, do the arithmetic on that, and then convert the answer back. The tablets show the working, along with the multiplication tables that supported it [FN5].

That middle step is the invention.

The number stopped belonging to the thing being measured. It could be lifted out, worked on, and put back. 5 no longer had to mean 5 sheep or 5 jars or 5 measures of grain. There was just 5.

It’s all about the money:

That abstraction is what makes value possible, and value is what makes money possible.

A unit of account is a number that survives being pulled away from whatever it was counting. The moment you can say that 3 sheep and 12 measures of barley are each worth 90 of something else, you have stopped talking about sheep and barley and started talking about value. Both sides of that comparison have to be the same kind of number before the comparison means anything.

What made money possible was a removal: taking the commodity out of the number and leaving the number standing there by itself.

Coins turn up much later, in Lydia in the 7th century BCE, and they are the least interesting part of this. The hard bit was done 1,300 years earlier by clerks doing sums about grain.

Zero, and my Persia problem:

Zero took much longer, and this is where my vague memory turns out to be half a fact.

In our system, the position of a digit changes its value. The 2 in 25 means 20. Any system that works this way needs a way of marking an empty slot, or 25 and 205 become indistinguishable. Babylonian scribes had a placeholder by around 700 BCE, but they never used it at the end of a number and never treated it as a quantity in its own right. It marked a gap without being one.

Zero as an actual number belongs to India. Brahmagupta’s Brāhmasphuṭasiddhānta, written in 628 CE, lays out rules for adding, subtracting and multiplying with zero, and gets thoroughly stuck on dividing by it. There is an older manuscript with a dot used as a placeholder, but its date is genuinely disputed and I would not lean on it [FN6].

The oldest Indian zero nobody argues about is carved at Gwalior, dated 876 CE.

Then it travelled. It was written up in Arabic in Baghdad around 825 CE by a scholar from the Persian-speaking world and reached European merchants through Fibonacci’s Liber Abaci in 1202 [FN7]. Which is why I had Persia in my head. My half-memory had kept the courier and lost the sender. That is roughly what happens to every idea that crosses a language on its way to you.

Then we went backwards:

Something odd happens next. Having spent 5,000 years arriving at 1 flexible system that could handle anything, we started building special-purpose number systems again. This time the constraint was not the commodity. It was the machine.

Binary is much older than computers and was studied as pure mathematics for centuries before anyone had a use for it [FN8]. What made it matter was the arrival of a machine whose basic component had exactly 2 states. A computer’s fundamental building block is on or off.

In 1937 a young engineer called Claude Shannon showed, in his master’s thesis, that an existing branch of algebra using only the values 1 and 0 described electrical switching circuits precisely [FN9].

That is the entire reason computers count in 1 and 0. Early machines that tried decimal instead were harder to build and harder to reason about, and by 1945 the argument was over.

Hexadecimal, which exists for us:

Binary suits the machine and punishes the reader. The number 3,735,928,559 needs 32 binary digits, and nobody can hold that on a screen and think.

Hexadecimal is base 16. It runs 0 to 9 and then carries on with A, B, C, D, E and F, so a single character can hold 16 different values instead of 10.

4 binary digits also produce exactly 16 combinations. That match is the whole trick. Chop the binary into blocks of 4, swap each block for its single hex character, and those 32 digits come out as 8: DEADBEEF.

Nothing is really converted. The bits are regrouped into clumps a human can read. Hexadecimal is a notation for people, sitting on top of a notation for machines, and it counts nothing that binary was not already counting.

What the answer actually was:

Numbers were assembled, slowly, by people solving specific problems, and none of the systems ever really got replaced. They piled up.

• Base 60 tells you the time.

• Base 12 sells you eggs and measures your shelf in inches.

• Base 10 states your salary, for no better reason than the number of digits on your hands.

• Base 2 carries every one of those figures through every electronic device you own

• Base 16 is how the people who maintain those devices read what base 2 is doing.

The Uruk scribe needed a system per commodity because number had not yet been separated from things. We need a system per machine because number has now been separated so completely that we can bend it to whatever we are counting with.

The single most consequential idea in the history of finance is that 5 can exist without being 5 of anything. It is so completely absorbed that it feels like a fact about the world rather than something people built.

It took roughly 40,000 years from the first notched bone to get there, and every balance sheet, every price, every salary and every currency sits on top of it.

So here is the short answer I owe my son. Nobody invented numbers. People needed to count things for different reasons. What they then invented was the idea that a number could exist on its own.

Footnotes

[FN1]: The term was coined by Kaufman, Lord, Reese and Volkmann, “The Discrimination of Visual Number,” American Journal of Psychology, 1949.

[FN2]: Jean de Heinzelin, who excavated the Ishango bone in 1950, read arithmetic into it (Scientific American, 1962). Alexander Marshack read it as a lunar record in The Roots of Civilization (1972). Olivier Keller has argued against both in Aux origines de la géométrie (2004). On Lebombo, the “at least” matters: the bone is snapped at one end, so 29 is what survived rather than what was necessarily carved, which makes the popular claim that it tracked a 29-day moon cycle shakier than it usually sounds.

[FN3]: The merger explanation is associated with Otto Neugebauer, The Exact Sciences in Antiquity (2nd ed., 1957). The divisibility and finger-joint explanations are widely repeated but not attributable to a single decisive source.

[FN4]: Hans Nissen, Peter Damerow and Robert Englund, Archaic Bookkeeping: Early Writing and Techniques of Economic Administration in the Ancient Near East (University of Chicago Press, 1993). The exact count varies between about 12 and 15 depending on whether sub-systems are counted separately. Denise Schmandt-Besserat’s related argument in Before Writing (1992), that clay tokens evolved directly into written numerals, is influential and contested; see Paul Zimansky’s review in Journal of Field Archaeology, 1993.

[FN5]: Nissen, Damerow and Englund, as above. This is the thinnest link in the chain, and worth saying so: the shift was gradual, and the tablets show a working method rather than a moment of invention.

[FN6]: The Bakhshali manuscript. A 2017 radiocarbon study by the Bodleian Library proposed that parts date to the 3rd or 4th century CE. Kim Plofker and colleagues challenged that reading in History of Science in South Asia the same year.

[FN7]: The Arabic account was by al-Khwārizmī, who came from Khwarazm and worked in Baghdad. Fibonacci’s Liber Abaci (1202) carried the system into European trade, where merchants took it up considerably faster than universities did.

[FN8]: A Sanskrit treatise on poetic metre, Pingala’s Chandaḥśāstra, sorts syllables into light and heavy in a structurally binary way; it is dated somewhere between 300 and 200 BCE. Leibniz published full base 2 arithmetic in Explication de l’Arithmétique Binaire (1703), attaching religious meaning to it and noting its resemblance to the hexagrams of the I Ching.

[FN9]: Claude Shannon, “A Symbolic Analysis of Relay and Switching Circuits,” MIT master’s thesis, 1937, published 1938. The algebra was George Boole’s, set out in An Investigation of the Laws of Thought (1854). Konrad Zuse’s Z3 (1941) ran on binary; ENIAC ran on decimal. John von Neumann’s First Draft of a Report on the EDVAC (1945) made the case for binary in machine design.

Leave a comment

Share

Subscribe now

His name was Kautilya, also called Chanakya, and he was the strategist behind Chandragupta Maurya, founder of the empire that first united most of the subcontinent. His treatise, the Arthashastra, traditionally dated to around 300 BCE, is a manual of statecraft, and a startling amount of it is about money. How to raise it, how to account for it, how to stop your own officials from stealing it, and above all, how to tax without strangling the thing you are taxing.

Subscribe now

The one-sixth share, adjusted for the soil

The heart of the Mauryan tax system was the bhaga, the king’s share of agricultural produce. The standard rate was one-sixth, the shadbhaga, roughly 16.67% of what the land yielded. But it was not a flat one-sixth applied blindly. The rate flexed with the quality of the land and the irrigation available, so better-watered, more productive land could be assessed differently from marginal ground. In an emergency, war or famine, the state could raise its demand, pushing the land levy as high as a quarter through a special exaction. The principle was already there in 300 BCE: tax according to capacity and lean harder only when the survival of the state required it.

Around that core sat a whole vocabulary of levies, each with its own name. Sulka was customs duty, split into an import duty and an export duty. Kara was a tax paid in cash. Vartani was a road cess on goods in transit.

There were taxes on mines, on liquor, on gambling, on entertainers, on the produce of forests. The state ran granaries, monitored markets, standardised weights, and stationed superintendents over every taxable activity it could identify. It was, on paper, one of the most thoroughly administered economies of the ancient world. What lifts the Arthashastra above a mere list is that Kautilya understood the danger in his own system. He warned, repeatedly, against taxing too heavily.

He grasped that a rate set too high shrinks the very activity it taxes, so the greedy king ends up collecting less than the patient one.

Modern economists, several thousand years later, gave this idea a name and a graph and called it the Laffer curve. The graph came in 1974, when an American economist named Arthur Laffer sketched it during a Washington dinner with 2 officials from the Ford administration, arguing against a proposed tax increase. The shape is simple enough to draw on a napkin, which is exactly where the story says it was drawn.

At a tax rate of 0%, the state collects nothing, for obvious reasons. At a rate of 100% it also collects nothing, because nobody works to hand over everything.

Between those 2 zeroes the revenue line has to rise and then fall, so a rate exists beyond which raising it further collects less. Where that peak actually sits is contested, and economists have argued about it for 50 years without agreement, which is why the curve gets invoked far more confidently in politics than in economics. What is not contested is the shape. India would go on to prove the falling half of it in the most expensive way available.

Kautilya had the insight without the graph. He put it in an image instead: a king should collect taxes the way a bee draws nectar from a flower, taking only what it can without harming the bloom. That sentence has outlived every empire that ignored it.

The Mughal machine: measuring the land

Fast forward 1,800 years, and India’s tax administration reaches its pre-modern peak under the Mughals, specifically under Akbar and the finance minister whose name every Indian schoolchild learns, Raja Todar Mal.

The problem Todar Mal solved was the oldest one in land taxation: how to set a fair, stable demand when harvests swing wildly from year to year. Tax too hard in a bad year and you ruin the farmer. Tax too softly in a good year and you starve the treasury. His answer, the dahsala or zabt system introduced in 1580, was to stop guessing and start measuring.

Officials surveyed the land with a standardised unit, classified it by how continuously it could be cultivated, and then calculated the state’s demand from the average yields and prices of the previous 10 years. The state’s share worked out to roughly 1/3 of average produce, and it was increasingly payable in cash rather than grain.

The leap here is conceptual, not just administrative. Todar Mal built the demand on a multi-year average, which smoothed out the violence of any single harvest and gave both the farmer and the state something they had never reliably had: predictability. He was reaching for the same thing every tax authority reaches for, a number that is fair enough to be paid and stable enough to be budgeted. The system was good enough that it outlasted the emperor who commissioned it and shaped Indian land revenue for centuries.

From Kautilya’s one-sixth-adjusted-for-soil to Todar Mal’s ten-year-average-by-land-class, India has an almost unbroken 1,800-year record of trying to assess the land scientifically, fairly, and in a way the farmer could survive. The instinct to measure before you tax is very, very old here.

The tax that built an empire, and the tax that broke it

Then the nature of taxation in India changed, because the people setting the rates were no longer accountable to the people paying them.

Under British rule, taxation became the machinery of empire, an extraction system whose purpose was to fund a foreign administration and, for a long stretch, to profit a foreign company. Land revenue was pushed hard, sometimes catastrophically. And one ancient, humble tax became the symbol of the whole arrangement.

Salt. The same salt that tax authorities had loved for 3,000 years, because everyone needs it and nobody can avoid it. The British salt tax made it illegal for Indians to produce their own salt from their own coastline, forcing them to buy it, taxed, from the colonial monopoly. Tax a person’s food and you anger them. Tax the single mineral that every human body requires, that the poorest labourer needs as much as the richest merchant, and you have created a grievance perfectly understood by everyone in the country at once.

Gandhi understood this with the precision of a man who knew exactly what he was doing. In 1930, he walked to the sea. The Salt March covered roughly 240 miles over 24 days, and at the end of it he bent down on a beach at Dandi and picked up a lump of natural salt, breaking the law in the simplest possible way. He did not attack the land revenue or the income tax or the customs system, though he could have. He chose salt, because salt was the tax that every single Indian paid and every single Indian could feel.

The most ancient and avoidable-proof tax in history became the most powerful piece of political theatre in the freedom movement.

It is the perfect closing of a loop that opened in Part 1. The salt tax was effective for exactly the reason Uruk’s grain levy was effective: it fell on the visible and the unavoidable. The colonial state used that ancient efficiency as an instrument of control. And a man in a homespun cloth used that same efficiency as the instrument of resistance.

When India became free, it inherited all of this: a deep tradition of careful assessment, a colonial habit of heavy extraction, and a brand-new power to decide for itself how a free people should tax itself. What it did with that power is the strangest and most entertaining part of the whole story, because it is the part you have actually lived through.

Next, Part 4: the 97.75% tax rate, the taxes that came and went, and the long road to GST. The Indian tax story, in your own lifetime.

Series Guide - History of Taxes:

• Part 1: What the First Writing Was For

• Part 2: The Tax That War Built

• ✓ Part 3: Kautilya to Gandhi

• Part 4: The Most Expensive 100 Rupees in History

Sources:

• Kautilya, Arthashastra (R. Shamasastry translation, 1915), books on the superintendent of commerce, customs, and revenue; specific references to the bhaga share (2.15.3) and crown-land income (2.24)

• D. K. Rangnekar and others; NIPFP working paper “On the Manu-Kautilya Norms of Taxation,” on the one-sixth core rate and its interpretation; on Kautilya’s anticipation of the Laffer curve and the “bee from the flower” image

• Dahsala system, and standard Mughal histories, on Raja Todar Mal’s zabt/dahsala revenue reform of 1580: ten-year averaging, standardised land measurement, four-fold land classification, the roughly one-third state share, building on Sher Shah Suri’s earlier experiments

• Standard histories of the British salt tax and the Salt Satyagraha (Dandi March, 1930): approximately 240 miles over 24 days

• Standard accounts of the Laffer curve’s 1974 origin, including Jude Wanniski’s 1978 article in The Public Interest; New York Times, “This Is Not Arthur Laffer’s Famous Napkin” (13 October 2017), on the disputed provenance of the Smithsonian napkin

  ---

  Share

  Subscribe now

  Leave a comment

The most common negotiation mistake looks exactly like good manners.

They offered you a promotion. You said thank you. And somewhere inside that moment of gratitude and relief, the conversation about scope, package and everything you were leaving on the table ended before anyone had opened it. A promotion is documentary evidence that your organisation wants to keep you. That is leverage, and it has a short shelf life.

Subscribe now

Preparation beats confidence

You do not need to feel fearless. You need to walk in knowing your market rate, sourced from LinkedIn Salary Insights, Glassdoor, and at least 2 direct conversations with people doing the same job elsewhere. Your instinct about what you are worth is not data.

You also need your impact written down before the meeting, in the language your employer already uses: projects delivered, revenue influenced, cost saved, cycle time reduced, people developed. Numbers beat adjectives. Every time.

Do not go first

If asked what you expect, redirect once: “I would like to understand the full scope of the role and the compensation band before I put a number on the table.” If pressed, offer a range with your target at the bottom of it, because the negotiation will travel downward from wherever it starts and you want your floor to be their ceiling’s neighbour. The first number spoken anchors everything that follows. Let it be theirs.

Negotiate the whole package

Most people negotiate base salary and leave the rest of the document untouched. The components worth more over time are usually elsewhere:

• Variable pay percentage and how it is measured

• Equity allocation and vesting schedule

• Employer NPS contribution under Section 124(2)

• Learning and development budget

• Notice period

• Flexible working

• Title

• Reporting line.

An extra INR 10,000 on base is worth roughly INR 1,20,000 a year before tax. A restructure that routes 14% of basic into NPS, or an accelerated vesting cliff, can be worth considerably more and often costs your employer less.

The gratuity conversation nobody has

If you are weighing an external move, work out where you sit against the 4 years plus 240 days threshold covered in Part 3. Cross it and your gratuity is payable. Fall short and it is not.

At a basic of INR 50,000 a month, the difference between leaving just short and leaving just past that line is about INR 1,15,000, rising to INR 1,44,231 at 5 completed years. That is a real number and it belongs on the table: “Making this move before my gratuity vests costs me approximately X. I would like the joining package to reflect that.” Most candidates never raise it. Most companies accommodate it when asked. Unclaimed money stays where it is.

Frame it as business, then stop talking

“Based on the scope of this role, my track record, and the market rate for this position, I believe a salary of Z reflects the value I bring.” That is a business case. “I was hoping for a bit more” is a personal appeal, and personal appeals get sympathy rather than money. Then say nothing. The silence will feel like a minute. It will be about 7 seconds. Hold it anyway. Whoever fills it usually concedes ground.

What to do right now:

• Research your market rate before any compensation conversation. LinkedIn Salary Insights, Glassdoor, and 2 peer conversations at minimum.

• Write down your impact in measurable terms. One page. Numbers in it.

• List every negotiable component, not just base. Variable, equity, employer NPS, L&D budget, notice period, flexibility, title.

• If you are considering an external move, calculate your gratuity position against the 4 years plus 240 days threshold, and put the number in the conversation.

• Rehearse the ask once, out loud. Deliver it. Then hold the silence.

  ---

  Leave a comment

  Subscribe now

  Share

I have a confession to make.

I did not finish a single craft project in school. Not one ever. There were half-made things, barely-started things, things that were ambitious for approximately forty minutes and then abandoned when something more interesting presented itself.

Nobody was harmed. The world continued turning. The incomplete embroidered pillow cover (VERY convent school coded) and the unfinished clay flowers and whatever else accumulated in that particular graveyard of good intentions cost me nothing except perhaps a few marks in a subject I was not losing sleep over.

I tell you this because the Malibadger was built, not born. And the building took time and specific experiences that made certain truths impossible to unknow once you had seen them clearly. This is one of those truths.

I have stood outside exam halls and watched people cry.

Several times, across different years, different institutions, different subjects. And the thing that strikes me every single time, the thing that has never lost its particular sting no matter how many times I witness it, is the specific quality of that regret.

If only I had studied that chapter. Two words. If only.

The most futile words in any language because the window has closed. The chapter that would have cost an hour to study properly is now costing everything, the grades, the opportunity, the version of the future that was sitting just on the other side of that unread page. The tab was left open. The energy was spent carrying it anyway. And then the thing it was pointing toward became permanently unavailable.

We don’t regret the things we finished badly nearly as much as the things we didn’t finish at all. A bad finish at least closes the loop. The incomplete thing stays open indefinitely, draining a little energy every day, and then one day the window closes and the open tab becomes a wound.

There was a man I heard speak once. I no longer remember the occasion or even the subject. What I remember is the quality of his thinking as it arrived in language.

His paragraphs had no loose threads. Each one started, developed, and ended. Complete, contained, no wasted words, no ideas left trailing, no thought begun and then abandoned mid-flight in favor of something shinier.

I remember thinking that this was a kind of integrity. Not just intellectual but moral. The respect for an idea that says I will see this through to its conclusion. I will not pick it up and put it down when it becomes difficult or when another idea arrives that seems more interesting. I will stay with it until it is complete.

That quality, I realised, was not just about speaking or writing. It was a way of being in the world.

A relationship with your own commitments that said when I begin something it deserves to be finished, not because finishing is always easy or rewarding or even particularly satisfying, but because the alternative is a life full of half-made things that never became what they were supposed to be.

Then there was a boss.

The standards were, by any reasonable assessment, impossible. Timelines that assumed a version of the working day that didn’t exist. Expectations that required you to find capacity you weren’t sure you had. The kind of demands that make you understand very clearly that comfort and completion are not always the same thing.

And yet. We met them. Not always perfectly. With considerable effort and occasional despair.

Because the alternative, the half-done deliverable, the almost-finished report, the presentation that was mostly there, was not actually an alternative at all. It was just a more exhausting version of not finishing, dressed up as progress.

What that boss understood, and what those impossible standards taught, was that the energy spent on an incomplete thing is never recovered. You pay the cost of starting. You carry the weight of the open loop. And then you don’t get the return that finishing provides, the closure, the momentum, the particular satisfaction of something that is done and can be set down.

Completing the thing, even the hard thing, even the thing that takes more than you expected, is almost always less expensive than not completing it. The energy economy of finishing is surprisingly generous.

And then there is motherhood.

Which is, at its most fundamental level, the longest and most demanding exercise in finishing what you start that a human being can undertake.

Your body spends nine months creating another person entirely. It is exhausting in ways that don’t have adequate language. And then the child arrives and the exhaustion doesn’t end, it simply changes form. It becomes the exhaustion of feeding and soothing and worrying and watching and being needed in ways that don’t pause for your own requirements.

Your body wants rest. It says so loudly and continuously. But you have created a life. And creating a life is not a task you can leave half done because you ran out of energy or motivation or because something more interesting presented itself.

Pick up the thing. Do the thing. Put it down. Move on.

That sequence, which sounds almost comically simple, is how it actually works. Not in grand gestures or heroic marathons of effort. In small, repeated acts of completion. The feed that needs to happen. The story that needs to be read. The school project (oh GOD the school projects!) that needs to be finished, which is perhaps the universe’s most pointed joke given where we started.

One thing at a time. Picked up, done, put down. Loop closed.

Over fifteen years, those small completions accumulate into something that looks, from the outside, like a raised human being. From the inside it feels like ten thousand individual decisions to finish the thing in front of you before rushing, breathless, to the next one.

I think about the craft projects sometimes. The abandoned pillow cover. The embroidery that never became anything (but I learnt the satin stitch!).

I don’t regret them. The lessons were elsewhere at the time and the stakes were appropriately low for a child still figuring out which things deserved her full attention.

But I understand now what they were the beginning of. The long education in why the loop needs to close. The exam hall tears that made the cost of the open tab undeniable. The speaker whose complete paragraphs showed me what intellectual integrity looked like in practice. The boss whose impossible standards turned out to be possible. The child who needed to be raised one completed act at a time.

Finish what you start. Not because finishing is always rewarding or easy or likely to produce the outcome you were hoping for.

Because the alternative is standing outside an exam hall with two entirely futile words and a window that has permanently closed.

Close the loop. It costs less than carrying it.

Subscribe now

Share

The man who changed it was William Pitt the Younger, and he did not do it because he had a theory about progressive taxation. He did it because he was desperate. By the late 1790s Britain had been at war with Revolutionary France for years, the Royal Navy was ruinously expensive, and borrowing alone could no longer cover the bill. In his budget of December 1798, Pitt reached for something no modern state had tried: a direct tax on what people earned.

Subscribe now

Levied from 1799, it worked on a sliding scale. Incomes below £60 a year paid nothing. Above that, the rate climbed with income, reaching a top rate of 10% on incomes over £200. That structure contains the entire DNA of the income tax you pay today: an exemption at the bottom, a rate that rises as you earn more, the whole thing assessed on the individual. Pitt hoped it would raise £10 million. It brought in around £6 million, which tells you that even the inventor of income tax discovered, immediately, that people are extraordinarily good at not declaring it.

Born of war, and dying with peace, twice

The thing to understand about the first income tax is that nobody, including the government, thought it should be permanent. It was an emergency measure, sold to a furious public as a temporary sacrifice for a national crisis. And the moment the crisis paused, so did the tax.

When Britain and France signed the Peace of Amiens in 1802, the income tax was abolished. War resumed within a year, the tax came back in 1803, and it was finally killed off in 1816, the year after Waterloo, to what one account of the time called a thundering cheer.

The British Parliament was so eager to be rid of it that it ordered every record of the tax destroyed, a gesture somewhat undermined by the duplicates already sitting safely with the King’s Remembrancer [the Exchequer official who kept the court’s records].

Income tax in Britain was born as a wartime expedient, one the public tolerated through gritted teeth and threw off the instant the guns fell silent. That it later crept back permanently, under Robert Peel in 1842, is one of the sharper ironies of fiscal history. The temporary tax became the most reliable revenue any modern state has ever known.

The pattern that began with Pitt repeated everywhere. The United States introduced its first income tax in 1861 to pay for the Civil War, then repealed it once the war was paid for. War created the desperation that overrode the public’s hatred of being taxed on their earnings, and war created something else the income tax needed: the administrative reach to assess and collect it. You cannot tax income until the state is organised enough to measure income. War forced states to get that organised.

No taxation without representation

While the modern income tax was being born of war, an older argument about tax was busy building the modern political world.

The principle had been forming for centuries in those scattered tax revolts: if you want my money, you must ask for my consent. In the second half of the 18th century it became a battle cry.

When the British Parliament tried to raise revenue from its American colonies through a series of duties and stamp taxes, the colonists’ objection was not only that the taxes hurt. It was that they had no representation in the Parliament imposing them. “No taxation without representation“ was a statement about tax that became a demand for political power, and it helped light the fuse of a revolution.

Across the Channel, tax sat near the centre of the French Revolution too. The pre-revolutionary system was loathed not just for its weight but for its unfairness. The nobility and the clergy, who owned the most, were largely exempt, while the burden fell on those least able to carry it. A tax system that exempts the powerful and crushes the poor makes a statement about who matters, and the French read it clearly. When the old order fell, the question of who should be taxed, and how fairly, was right at the heart of what replaced it.

This is the second great theme of tax history, sitting alongside the first. Tax is the thread that ties the citizen to the state. Stretch it too far, or tie it too unevenly, and it snaps, and when it snaps it can take a government down with it.

The state learns to tax what you buy, all over again

The 19th and 20th centuries built the tax state we now live inside. Income tax spread and deepened, and broad-based taxes on earnings became the backbone of government revenue across the industrialised world. The invisible had been made visible, through paperwork, employers, banks and eventually computers, until the state could finally see income clearly enough to take a steady share of it.

But the oldest instinct in taxation, to tax the thing you can see at the point of sale, never went away. It came back in the 20th century in a more elegant form than any ancient salt tax.

This is where the tax world splits into 2 halves, and both halves survive today for good reason. A direct tax is levied on a person and paid by that person: income tax, wealth tax, the levy that follows you. An indirect tax is levied on a transaction, collected by whoever sells you the thing, and buried in the price you pay: customs duty, excise, GST.

The state keeps both because each covers the other’s weakness.

Direct taxes can be made to rise with capacity, so someone earning 10 times more can be asked for a larger share, which is the only way a tax system can be fair on purpose. But they only reach the people the state can see earning, and in most countries that is a minority of the population.

Indirect taxes reach everybody who buys anything, including every earner the income tax never finds, and they arrive steadily, because people keep eating and washing and travelling through recessions. They are also blunter. The same tax on a bar of soap takes a far bigger bite out of a labourer’s day than a banker’s, which is why modern systems exempt or lightly tax essentials and load the higher rates onto things nobody needs.

Fairness underscores direct taxes. Reach and reliability are supplied by indirect taxes. Running both is how a treasury gets a system that can be aimed at capacity and still keeps paying when the economy turns. That is the whole reason your salary slip and your grocery bill are both taxed, in different ways, by the same government.

In 1954, France implemented the first modern value-added tax, designed by a tax official named Maurice Lauré. The idea, which had been floating since a German industrialist sketched it in 1918, was clever. Instead of taxing the full price of a good only at the final sale, you tax only the value added at each stage of production, with businesses claiming credit for the tax already paid by their suppliers. It collected reliably, it was hard to evade because every business had a paper trail interest in the stage before it, and it raised enormous sums without anyone having to declare their income.

VAT was the ancient idea reborn with modern accounting. Tax the transaction, not the person. It spread across Europe and then the world with remarkable speed, and by the time it reached India, decades later, it would arrive under a different name and trigger the largest tax reform in the country’s history. That is Part 4.

But to understand India’s tax story, you have to go back much further than 1954, or 1799, or even Rome. Because while Europe was farming out tax collection to the publicani, an Indian treatise was already laying out a tax system of breathtaking detail, and an Indian empire was already running it.

Next, Part 3: Kautilya’s astonishing tax code, the Mughal revenue machine, and the British salt tax that Gandhi walked to the sea to break.

Series Guide - History of Taxes

• Part 1: What the First Writing Was For

• ✓ Part 2: The Tax That War Built

• Part 3: Kautilya to Gandhi

• Part 4: The Most Expensive 100 Rupees in History

  Subscribe now

  Share

Leave a comment

Sources:

• History of Taxation in the United Kingdom, and History Hit, on Pitt the Younger’s income tax: December 1798 budget, levied 1799, top rate 10% on incomes over £200, exemption below £60, £6 million raised against a £10 million target; abolition 1802, reintroduction 1803, abolition 1816 and the ordered destruction of the tax records; reintroduction by Peel 1842

• Brewer Morris and standard fiscal histories, on the first US income tax (1861, Civil War) and its repeal

• Standard accounts of “no taxation without representation” and the fiscal causes of the American and French Revolutions

• Value-Added Tax, on Maurice Lauré and the first modern VAT (France, 1954), the von Siemens concept (1918), and the global spread of VAT

Walk into an electronics shop in any Indian city this month and you will feel the good half of this story before you understand it. The air conditioner is cheaper than it was a year ago. So is the phone, the power bank, the solar inverter on the shelf behind the counter. Your money goes further than it used to. You leave happy.

Subscribe now

Now cross town to a small factory that makes one of those same things. The owner is not happy. He is watching the same cheap goods that delighted you land at his warehouse door at a price he cannot match, priced below what it costs him to make them, and he is doing the arithmetic on how many more months he can keep the line running.

One of you is getting richer and one of you is being pushed out of business, and you are both reacting to the identical thing: a wave of goods arriving from China at prices that should not be possible.

To understand how the shopper’s windfall and the factory owner’s ruin are the same event, you have to start with a man who died 3 days ago, in Beijing, whose name most Indians have never heard.

Zhu Rongji died on 12 August 2026, aged 97. He may be the most consequential and unknown economic figure of the last half-century, and the timing of his passing is grimly apt, because the crisis now spilling out of China is the precise thing he spent his career preventing.

Zhu was China’s premier from 1998 to 2003, an acerbic, engineer-trained technocrat known at home as the economic czar. He did the hardest thing a Chinese leader has ever done to his own economy. He closed things down.

His programme to restructure China’s bloated state-owned enterprises threw an estimated 30 million people out of work, because the factories they worked in were making things nobody wanted at prices that made no sense, and Zhu decided the country could not afford to keep them alive.

He forced the discipline that a market is supposed to impose on its own: when a factory loses money, it shuts, and the people and capital move to something that does not. He then negotiated China into the World Trade Organization, the act that turned the country into the workshop of the world.

Almost everything China has become, it became on the foundations Zhu poured. And the crisis it is now exporting to the rest of the world, the one landing on that Indian factory owner’s doorstep, is what happens when his core principle, that a factory which chronically loses money cannot simply be kept alive forever, gets abandoned.

What overproduction actually means

A country overproduces when its factories can make far more than its own people are able or willing to buy, and far more than the world needs on top of that. The excess has to go somewhere. It can’t sit in warehouses forever so it goes abroad, sold cheap, often below what it cost to make, because a sale at a loss is still better than a product that never sells at all.

China is doing this now across a startling range of goods. Not just the cheap, labour-heavy things it has always made, but cars, solar panels, batteries, steel, cement, chemicals, and the legacy chips that run ordinary electronics. Growing inventories at home have pushed Chinese firms to redirect the surplus abroad, and the low prices travel with the goods.

China exported USD 8.6 billion of cars in 2019 and USD 77.6 billion of them by 2023, a 9-fold jump in 4 years. The world did not start needing that many additional Chinese cars. China simply started making far more than it could sell at home.

This is not new information. Economists have 2 names for what is happening.

1. The first is that China is “exporting deflation”: When cheap Chinese goods flood a market, they drag down prices there, so China effectively ships its own falling-price problem to everyone else.

2. The second, borrowed from a famous piece of economics, is “China Shock 2.0”: The original China Shock was the wave of cheap Chinese goods that hollowed out Western factory towns after China joined the world trading system in 2001. The sequel is the same shock arriving again, this time in higher-value goods like cars and clean-energy equipment, and this time landing on developing economies as much as rich ones.

Zhu’s rule existed to prevent this exact situation. A market economy is supposed to punish overproduction.

When solar factories churn out more panels than the world will buy, prices fall, the weakest producers go bankrupt, capacity leaves the industry, and the survivors return to making a profit. That painful correction is the mechanism that keeps supply and demand roughly in line. China has switched it off.

Why China cannot stop

In a market economy, an unprofitable factory closes because no one will keep funding its losses. In today’s China, it stays open, because the state keeps the money flowing to it through cheap loans and local-government support, and because the ruling ideology treats productive capacity as national strength rather than as a business that has to earn its keep. A factory is jobs, and jobs are stability, and capacity is power. So, the losing factory is kept alive, and it keeps producing, and the surplus keeps growing.

Take Chinese solar panels. Chinese solar manufacturers recorded losses in 2024 and 2025 while exporting at record levels. A rational market would have shut the loss-making capacity. China expanded it, because investment in manufacturing has kept growing at double digits even as consumption stays weak.

Factories are not being built because of additional demand. They are being built because building factories is what the system rewards.

The buyers who never came

If China makes so much, why can its own people not simply buy it, and soak up the surplus at home? The answer is a piece of economic theory. It’s called precautionary saving, an economist’s term for the most human thing in the world.

When there is no real pension, no free healthcare, no reliable safety net, a rational person hoards money against every disaster they can imagine, because the state will not catch them if they fall. A generation of Chinese was taught, correctly, that the only protection was the balance in their own account, so they built that balance and refused to touch it. They saved against a future they did not trust.

Fear, not greed, drives it.

At the level of the individual household, Indians have historically saved about as much of their income as Chinese households, sometimes more. Where China pulls far ahead is the national savings rate, savings as a share of the whole economy, which has run near 43% of output against India’s 30%.

The gap is not really about thriftier Chinese families. It is that in China the state and the corporations save on a colossal scale too, ploughing retained profits and government surpluses back into yet more factories and capacity rather than into wages and consumption. The whole system, household and state and firm together, is tuned to save and build rather than to spend.

Zhu built the productive half; the factories and the access to world markets. The other half, the pensions and healthcare and steadily rising wages that would let people spend freely on what those factories make, was never built to match it, and his successors chose to pour money into yet more production rather than construct it.

All that saving turns from a private virtue into a collective trap, through the idea John Maynard Keynes called the paradox of thrift.

Saving is sensible for one household. It is ruinous when a whole economy does it at once, because your spending is my income and my spending is yours, so when everyone holds back on the spending, all of the incomes fall together, and the economy seizes.

A country of relentless savers, each behaving wisely alone, can starve its own factories of the customers they were built to serve. That is precisely China’s condition. It saved, at every level, as it was built to, and left the shelves full and the buyers absent. The surplus had nowhere to go but out.

The overproduction making its way abroad and the under-consumption at home is one problem seen from two ends. The factories cannot stop making, because the system rewards making. The households cannot start buying, because fear rewards saving. What cannot be sold at home is loaded onto a ship.

Zhu’s whole reputation rested on being willing to do the opposite: to let the 30 million go, to close what could not pay. Understanding why his successors will not do the same is the key to the whole crisis, because the shift was neither immediate nor inevitable.

Why the rule was abandoned

The reversal did not arrive with Xi Jinping. In November 2013, a year into his leadership, the Third Plenum of the 18th Party Congress declared that the market would play the “decisive role in allocating resources,” in a reform blueprint released under Xi’s own authorship. For a moment it looked as though Xi might deepen Zhu’s project rather than bury it. What turned him, by the reading of analysts who watch Beijing closely, was a crisis.

In 2015, a stock market bubble that Beijing itself had encouraged, inflated by ordinary investors buying shares with borrowed money, burst in the middle of the year.

The state did not let the market find its floor. It intervened with everything it had: halting trading in hundreds of companies, banning big shareholders from selling, and ordering state-backed funds to buy shares directly to prop up prices. It fell anyway. Much of the rally had been built on borrowed money, and once prices fell, investors who had bought shares with loans were forced to sell to cover them, which pushed prices lower, which forced the next round of selling, a mechanical spiral that no official cheerleading could halt.

The market lost close to half its value from its peak over the months that followed, and the sight of the state deploying its full arsenal and failing anyway did as much damage to confidence as the fall itself. A market economist would have drawn one lesson from that: do not inflate bubbles, and do not fight corrections.

Kevin Rudd, the former Australian prime minister and a serious China scholar, argues that Xi drew almost the opposite lesson. The episode hardened Xi’s underlying scepticism about trusting markets over the instruments of the Party and the state. On this reading, 2015 did not teach Xi that state meddling had made things worse. It taught him that markets left to themselves produce chaos the Party is then forced to clean up, and that the safer course is more control, not less.

In Xi’s economic model, state ownership must play the leading role, and productive capacity is treated as a pillar of national security rather than as a business that has to earn a return. Once you classify a factory as security, closing it when it loses money stops looking like discipline and starts looking like disarmament. That is the reasoning that keeps the loss-making solar plant open. It is not a failed bet to be wound up. It is an asset of the state.

There are differing perspectives here among the experts. The dominant view holds that Xi represents a counter-reform era, a real departure from the direction Zhu set. A counter perspective, argued by the scholars Wendy Leutert and Sarah Eaton in The China Quarterly, holds that Xi’s handling of state enterprises is a deepening of trends already present under his predecessors rather than a clean break. Both camps agree on the outcome.

Loss-making capacity that a normal market would close, is being kept alive by the state. Zhu used the crises of his day, the internal reckoning with bankrupt state firms and the external demands of joining the WTO, as the lever to force painful market discipline. Xi faced his own crisis in 2015 and used it to justify the reverse.

Same country, crisis as the turning point in both cases, opposite conclusions drawn. The bill for Xi’s conclusion is now landing outside China’s borders.

How the surplus reaches the rest of the world

The surplus does not spread evenly. It follows the path of least resistance, and that routing matters most for India.

The United States has largely walled itself off. Tariffs and trade restrictions have made the American market hard for Chinese goods to enter, and the effect shows in the numbers: China’s exports to the US fell about 20% across 2025 as trade diverted elsewhere.

The goods that can no longer get into America look for somewhere more open to go, and they have been flowing instead toward Southeast Asia, Africa, Latin America, and the European Union. Those regions absorbed the bulk of China’s export growth in 2025. Europe has felt the pressure sharply enough that, by one industry estimate, the bloc has been losing up to 500 manufacturing jobs a day to the surge.

India sits squarely in the path of the redirected flow. It is a large, growing, relatively open market next door to the source, and it is exactly the kind of destination surplus goes looking for when the richer markets close.

India has been raising anti-dumping duties on Chinese steel, chemicals, and other goods precisely because they are arriving cheap enough to undercut anything made here, and “dumping”, in trade law, means selling below the normal price, which is what a country with too much supply does to clear it.

Why this is India’s problem specifically

For a piece about someone else’s factories, this lands very close to home, and in opposite ways.

1. Cheap Chinese goods are, initially, good for the buyer. Cheaper solar panels help India electrify faster. Cheaper batteries and components lower the cost of everything built on them. A country trying to green its grid and equip its people can do it faster and cheaper when the inputs arrive underpriced from next door.

2. The gravest danger of a flood of cheap imports is that it drowns your own industry before it can stand. India’s whole current economic ambition rests on building a manufacturing base: making the electronics, the solar cells, the components it now imports, so that it makes more of what it consumes and earns more of what it spends. A domestic solar factory or chip plant, still finding its feet, cannot survive against Chinese product dumped below cost. The cheap import that helps the buyer today can kill the domestic maker who was supposed to supply that buyer tomorrow.

That is the trap. The short-term saving and the long-term ambition point in opposite directions, and the country has to choose which it is protecting.

This is already happening to real Indian companies, in real sectors, right now. The clearest battleground is chemicals. India’s Directorate General of Trade Remedies, the body that investigates dumping, has been defending domestic makers like SRF and Deepak Nitrite across a widening list of goods in the last 2 years:

• A 5-year duty on certain Chinese electrical steel in 2025

• A duty on the refrigerant gas R-134a where SRF is the only Indian producer

• Duties on a string of water-treatment and industrial chemicals

Each of these is a small official confirmation of the finding that goods were arriving at prices Indian factories could not match, and that without protection the domestic maker would have lost.

A wave of underpriced Chinese goods makes importing even more tempting and building even harder, because the maths for anyone trying to open a factory here gets worse with every cheap container that lands. The overproduction next door is a headwind pushing directly against the one thing India most needs to do.

Every anti-dumping duty is the government trying, one product at a time, to hold that headwind back long enough for something to get built behind it.

The counterarguments

Chinese officials argue (in defence of overproduction) that the world genuinely needs vastly more clean technology, and that Chinese capacity is helping to fight climate change and hold down prices everywhere. There is truth in it. Global demand for electric vehicles and clean energy is set to grow enormously, and some of what looks like excess capacity today may simply be early capacity for demand that has not arrived yet. Rising future demand does not erase a present glut, but it complicates the picture.

The feared global price crash has been slower and milder than the alarm suggests. The heavy deflationary pressure has so far been felt mostly inside China, not fully passed on to world prices, because Chinese firms have been absorbing the losses rather than cutting export prices to the bone. The real risk is those firms dropping export prices much further to break into markets or to punch through new tariff walls. The danger is less what has already happened to global prices and more what China’s idle capacity leaves it able to do next.

How this reaches your life

All of this can feel like a story about governments and economists and trade analysts and professors but it reaches your own life in 4 direct ways:

1. As a shopper, you gain. The cheap Chinese air conditioner, phone, or solar inverter genuinely costs you less than it would if India made all of it, and that lower price is a benefit to your household budget.

2. As someone who works or invests, it is more complicated. If you work in or hold shares in a sector exposed to Chinese dumping, chemicals, steel, solar, electronics, the same cheap imports that help you as a shopper are squeezing the company that pays you or that you own.

3. As a saver, watch the knock-on effect. If cheap imports stop India from building its own factories: the country keeps importing what it could make; the rupee stays weak; your savings keep losing value against a falling currency.

4. As a citizen, understand what the duties are for. Every time you read that India has imposed an anti-dumping duty on some Chinese chemical or steel product, that is the state buying time and space for a domestic industry to survive the glut of goods sold below cost.

Epilogue

Zhu Rongji built the machine that made China the workshop of the world, and he built into it a rule that kept the machine honest: capacity that loses money must be allowed to die.

The system that outlived him kept the machine and discarded the rule. The result is a country producing far more than it or the world can absorb, held up by a state that treats every factory as too strategic to close, sending the difference abroad at prices designed to clear inventory rather than earn a profit.

For India, the overproduction next door is both a discount and a threat. The cheap panel that helps electrify a village is made in the same factory whose exports could smother the Indian panel-maker before it scales. Which of those matters more is now a policy choice India is actively making, duty by duty, sector by sector.

Which returns us to the two people we started with, the shopper leaving happy and the factory owner doing his grim arithmetic. They were never in different stories.

The cheaper air conditioner and the closing factory are the opposite ends of the same chain that runs back through a shipload of unsold Chinese goods, through a nation of frightened savers who would not spend, through a state that will not let a losing factory die, to a man who understood that a factory which cannot pay its way is not strength but liability, and who is no longer there to say so.

Zhu understood, more clearly than almost anyone, that an economy is not strong because its factories are busy. It is strong when its factories make things worth making at prices that pay. The distance between those two ideas is the whole of the story now unfolding.

Leave a comment

Share

Subscribe now

Sources

1. Zhu Rongji death, 12 August 2026, Beijing, aged 97, of illness: Xinhua obituary as reported by CNN, SCMP, Britannica, 12 August 2026

2. Zhu Rongji restructuring of state-owned enterprises, estimated 30 million laid off (World Bank estimate; AFP reports 30 million over five years); WTO accession 2001; “grasping the large, letting go of the small”: Xinhua obituary and SCMP, Yahoo/AFP, Vision Times, 12 August 2026

3. China auto exports rising from 8.6 billion dollars (2019) to 77.6 billion dollars (2023): Deloitte Insights, Global Impacts of Chinese Overcapacity

4. Overcapacity spread across traditional and emerging sectors (steel, cement, glass, autos, solar, batteries, legacy chips); excess capacity sufficient to potentially double NEV and lithium battery exports: Coface, “Made in China”; MERICS overcapacity analysis

5. Chinese solar manufacturers recording losses in 2024-25 despite record exports; double-digit manufacturing investment growth amid weak consumption: MERICS overcapacity analysis

6. China’s exports to the US falling about 20% across 2025 to USD 419.5 billion; Africa, Southeast Asia and Latin America absorbing nearly three-quarters of China’s export growth; EU losing up to 500 manufacturing jobs per day: US-China Economic and Security Review Commission, China Bulletin, February 2026; MERICS

7. Overcapacity as deflationary pressure felt within China more than in global prices; firms absorbing losses rather than cutting export prices: Rhodium Group, “Overcapacity at the Gate”; Deloitte Insights

8. China’s official position (Xi Jinping on “no overcapacity”; clean-technology and climate argument; global EV demand projections): CKGSB Knowledge, China manufacturing overcapacity analysis

9. 2013 Third Plenum “decisive role of the market” and the reform blueprint under Xi’s authorship: Seafarer Funds, “Security Over Growth”; Carnegie Endowment, “A Chinese Puzzle”

10. 2015 stock market crash, state intervention, and its effect on Xi’s thinking: Kevin Rudd / Asia Society Policy Institute, “Xi Jinping’s Pivot to the State”

11. Xi-era SOE governance as departure versus deepening (the scholarly disagreement): Asia Society and Carnegie (counter-reform reading); Wendy Leutert and Sarah Eaton, “Deepening Not Departure,” The China Quarterly (2021), for the dissent

12. India anti-dumping duties on Chinese steel, chemicals and other goods: Government of India / Ministry of Finance notifications and DGTR recommendations, 2025-26, including five-year duty on cold-rolled non-oriented electrical steel (DGTR, September 2025) and duty on refrigerant R-134a; Deccan Herald and Economic Times reporting on the notifications

13. Chinese dumping pressure on Indian chemical makers (SRF, Deepak Nitrite, Atul): Systematix Research report as reported by ANI, November 2025; Business Standard and market coverage of SRF anti-dumping duty, December 2025

14. “Exporting deflation” framing: ING Think, “China’s overcapacity debate”, 2024, among others

15. “China Shock 2.0” framing: US-China Economic and Security Review Commission, 2025 Report to Congress and February 2026 bulletin; original “China Shock” literature (Autor, Dorn, Hanson)

16. Savings rates: China national/gross savings near 43% of GDP versus India near 30% (AMFI-Crisil Factbook 2024; CEIC); household-level comparison showing Indian household saving historically at or above Chinese household saving, with China’s national lead driven by corporate and government saving: CEPR / VoxEU, “Why do Chinese households save so much?

Prologue

Everything in this section is memory and rumour. None of it is sourced, none of it is evidence, and none of it supports any finding in the rest of this piece. It is here because the rest of this piece is about what was never written down, and this is what that felt like from inside a housing society in Mumbai. One item in it turns out to have a document attached, and I will show you that document later.

Subscribe now

Remember when we could not pay for things?

I remember an ordinary morning in November 2016, standing at a vegetable cart on a Wednesday morning holding a ₹500 note that had been money on Tuesday night, watching the seller shake his head sheepishly. The auto driver followed suit. The chemist accepted it on the first day and refused it on the fourth. Within a week the whole country had learned to open every transaction by asking whether the other person had change, in a tone that started hopeful and stopped being hopeful quite fast.

Then the lines. They formed before the bank branches opened and did not thin out till after they closed. People brought folding stools and umbrellas and water. Somebody’s driver held a place for 4 hours so their employer would not have to. The ATMs ran dry within hours of being filled and then stayed dry, because the machines had been built to dispense a note of one size and the new note was a different size, so every tray in the country had to be physically refitted before it could hand out the currency that existed.

We learned to pay by phone because the alternative was not paying at all. People who had spent years declining to install a banking application installed 3 in a fortnight. Shopkeepers who had never taken anything but cash, taped a printed QR code to the counter and learned, badly and then well, how to check whether the money had arrived.

The ones I think about are the seniors in my complex. Their pensions came in cash, their household money sat at home in an almirah because that is what a lifetime had taught them to do, and their children were at work or in another country. They did not know how to install and then use the new technology, and the people who could have taught them were standing in a different line. One by one they went down to the bank themselves, in the sun, with their papers in a plastic folder, to explain to a clerk 40 years younger where their own savings had come from. Some of them could not produce the answer in the form the counter wanted. The indignity of that has stayed with me longer than the inconvenience of any of it.

Underneath all of it ran a total absence of reliable information. The rules moved: one exchange limit on Thursday, a smaller one by the weekend, an exemption announced and then withdrawn, a deadline extended for one category and not another. Into that vacuum came the waves of social media forwards. The new ₹2000 note had a nano-GPS chip embedded in it that would let the government track hoarded cash from a satellite. Some banks would take old notes and some would not. A cousin’s colleague had heard from someone at the RBI.

And there was the lore, which everybody repeated and nobody could source. Truckloads of cash burning in fields somewhere in the north. Gunny sacks going into canals at night. Jewellers who kept their shutters half open past midnight on 8 November, backdating bills at whatever price the buyer wanted. A builder who paid a year of wages in advance to 200 labourers in a single afternoon.

And within about 2 weeks, everybody knew a man who knew a man. He would take your unusable notes and hand back usable ones, and keep 20% of them, or 30%, or 40%, priced according to how frightened you were. That last one is the item with a document attached.

What follows is only what was written down.

The Ledger Nobody Assembled

Every number required to evaluate demonetisation was published by the Government of India or the Reserve Bank of India. Each one is public, downloadable, and has been for years. They have never been printed on the same page.

This piece prints them on the same page. It does not tell you what conclusions to draw. It assembles a ledger from the record kept by the institutions that ran the exercise, and ends with what that record does not contain.

A note on method. Where this piece says that no outcome was published, it means that no such outcome was identified in the official sources listed at the end, or in the parliamentary, departmental and institutional records examined while assembling this ledger. “No published record” is a statement about that search, not a claim of universal knowledge.

3 findings can be counted from the record. Each is directly from government documents.

ONE - The reports that went nowhere:

When a bank sees a transaction that does not fit what it knows about a customer, it is legally obliged to file an alert with the Financial Intelligence Unit at the Ministry of Finance. The alert is called a Suspicious Transaction Report. It is the formal trigger that is supposed to start an inquiry.

Banks filed 361,214 Suspicious Transaction Reports in 2016-17. That is 5.9 times the 61,361 filed the year before. In the year of demonetisation, the banking system generated more than 1/ 3 of a million formal alerts that transactions did not add up.

What is not in the record is what happened to any of them. No published account identified for this piece traces the 361,214 reports to inquiries opened, cases assessed, prosecutions brought, or money recovered. The alerts were filed. Their disposal was never published.

TWO - The deaths that were, and were not, counted:

Three arms of the government were asked, at different times, how many people died in connection with demonetisation. Their three answers cannot all be true.

• 18 March 2017, Lok Sabha. The Minister of State for Finance, in a written reply, stated that no official report had been received on deaths that occurred during the course of currency exchange and withdrawals.

• 18 December 2018, Rajya Sabha. The Finance Minister, answering a question from Shri Elamaram Kareem, stated that the State Bank of India had reported 3 staff members and 1 customer as having died during the period of demonetisation. He stated that no other public sector bank had reported any such death, and that ₹44 lakh had been paid in compensation to the families.

• Right to Information reply. The Prime Minister’s Office, asked the same question, stated that it held no information on the number of deaths.

The finding is the contradiction. The government paid ₹44 lakh in compensation for deaths during demonetisation, and separately told Parliament that no report of any such death had been received, and separately told a citizen that it held no information at all. Compensation was paid for an event the record elsewhere says was never recorded. No consolidated count of demonetisation-linked deaths was identified in the official record examined for this piece.

THREE - The man with the document:

The last rumour in the prologue was the man who would take your dead notes and hand back live ones for a cut. That one is in the tax department’s own files.

In December 2016 the Central Board of Direct Taxes described a decoy operation run by its Investigation Directorate in Mumbai, in a statement reported on 9 December 2016. Officers posing as customers approached a syndicate that was converting banned notes into new currency for other people. The syndicate’s mediator agreed to the exchange for a commission of 35%, was caught in the act, and new notes of ₹29.5 lakh were seized. The rate the whole city had whispered about appears in an official account of a sting, quoted and agreed at 35%.

It was not only street operators. The Income Tax Department’s Operation Clean Money Strategic Plan, the department’s own published planning document, records a search on the head office and branches of a Delhi cooperative bank that, in the document’s words, used demonetisation as an opportunity to earn illegal commissions for handling unexplained cash.

The mechanism was benami accounts, meaning accounts opened in a name that is not the true owner’s. The document records that the chairman and his associates opened more than 1,200 new accounts and reactivated 600 dormant ones after 8 November 2016, many without KYC (Know Your Customer identity) documents, that more than ₹120 crore was deposited by 26 December, and that most of it was routed back out to the beneficiaries.

The document proves the racket existed and priced itself. What the document does not contain is what happened next. No published record identified for this piece shows the syndicate prosecuted, the bank’s management convicted, or the ₹120 crore recovered. The rumour got a document. The document, like everything else in this piece, stops before the outcome.

The rest of the black hole:

The 3 findings above are not isolated. They are the ones that can be counted cleanly. Around them sits a wider set of numbers that entered the public record and then stopped before any outcome was published.

• 17.92 lakh persons were identified for verification under Operation Clean Money in 2017, after their cash deposits did not match their tax profiles. About 3.78 lakh cases were retained for closer examination. No published account records how many explained their deposits, how many were assessed, or how much tax was finally collected.

• 2.24 lakh companies were struck off the register. No published account records what was recovered from them.

• More than 400 cases were referred to the Enforcement Directorate and the Central Bureau of Investigation. No published account records their disposal.

• Terror financing was named as an objective on 8 November 2016. No assessment of demonetisation’s effect on it was identified in the official sources examined for this piece.

• The impact on industry and employment was never studied. Answering in the Rajya Sabha on 18 December 2018, the Finance Minister stated that no specific study had been done by the government on the subject. This is not a missing document. It is an official confirmation that the measurement was never made.

• Household consumption for the year after demonetisation was not published. The National Statistical Office ran its 75th round consumer expenditure survey from July 2017 to June 2018. On 15 November 2019 the Ministry of Statistics and Programme Implementation announced it would not release the results, citing data quality. The next comparable survey was conducted in 2022-23. The year that followed demonetisation has no released consumption data

Epilogue:

The state remembered how to count when it wanted to. It counted 361,214 suspicious transactions in a single year. It counted 35% off the top of a syndicate’s exchange rate, down to the last ₹29.5 lakh seized. It counted ₹44 lakh in compensation and paid it to the families.

Then, on every question that would have told us whether any of this worked, the counting stopped.

Almost 10 years on, the ledger is still open, and the entries that would close it were never made.

Subscribe now

Share

Leave a comment

Primary Sources

1. Reserve Bank of India, Annual Report 2016-17, Chapter VIII (Currency Management), Table VIII.9: Suspicious Transaction Reports filed with the Financial Intelligence Unit.

2. Government of India, “A Year After Demonetisation”, 8 November 2017: Operation Clean Money figures, companies struck off, cases referred.

3. Central Board of Direct Taxes / Income Tax Department, Operation Clean Money statements, January to April 2017.

4. Lok Sabha, written reply of the Minister of State for Finance, 18 March 2017: no official report of deaths received.

5. Rajya Sabha, reply of the Finance Minister to Shri Elamaram Kareem, 18 December 2018: SBI-reported deaths and ₹44 lakh compensation; no study on industry and employment.

6. Prime Minister’s Office, reply to Right to Information application on demonetisation-linked deaths.

7. Ministry of Statistics and Programme Implementation, statement of 15 November 2019 on the Consumer Expenditure Survey 2017-18.

8. Income Tax Department / Central Board of Direct Taxes, Operation Clean Money Strategic Plan: search on a Delhi cooperative bank; illegal commissions, benami accounts, ₹120 crore deposited and routed back.

9. Central Board of Direct Taxes statement, reported 9 December 2016, on the Investigation Directorate’s Mumbai decoy operation: 35% commission agreed, ₹29.5 lakh in new notes seized (via contemporaneous reporting of the CBDT statement).

   ---

The oldest writing anyone has ever found is not a prayer, a law, or a love letter. It is an account. Around 3300 BCE, in the Sumerian city of Uruk (what is now southern Iraq), scribes pressed marks into wet clay to record quantities of grain, heads of cattle, jars of oil, and days of labour owed. For its first several centuries, writing did almost nothing else. The marks counted things. They tallied what came in, what went out, and what was due. A reader 5,000 years later can recognise the format instantly, because it is a ledger.

Subscribe now

That fact reorders something most of us thought we knew. We tend to imagine writing arriving so humans could tell stories or speak to gods. That was for the leisure hours. Writing first arrived so a state could keep track of what people owed it.

The tablets grew out of an even older habit: little clay tokens, sealed in clay envelopes called bullae, used to record transactions before anyone could write at all. Writing was the upgrade. The thing being recorded was, from the very beginning, a tax.

For almost the whole of human history, a state could see what you grew, what you carried through its gates, and what you sold in its markets. It could not see what you earned. So, it taxed the things it could count. The tax on income, the one that eats a slice of your salary before you have touched it, is a startlingly recent invention. The arc from “tax what you can see” to “tax what you make” is the story of these four pieces. It begins here, with grain.

The thing a state can count

An empire cannot read minds, but it can stand at a gate.

This single limitation explains most of ancient and medieval taxation. Income is invisible. A farmer’s harvest is not. Goods moving along a road can be stopped and assessed. A market stall is a fixed place with a measurable turnover. Salt comes from a few known sources and everyone needs it. So, the earliest tax systems clustered, sensibly, around the visible and the unavoidable: land and its produce, trade and its routes, specific goods at the point of sale.

Egypt taxed grain and labour, and sent scribe-assessors out to measure the standing crop before harvest so the state’s share could be calculated in advance. The annual Nile flood was watched closely, because a high flood meant a good harvest which meant a larger assessment. Tax was tied to the river. The reason was simple administrative honesty: the produce of the land was the one form of wealth a Pharaoh’s officials could actually see and count.

The most famous object in the history of writing is, underneath, a tax document. The Rosetta Stone, the slab that let scholars finally read Egyptian hieroglyphs, carries a decree issued by a council of priests at Memphis in 196 BCE. Strip away the romance of decipherment and read what it actually says, and a large part of it is fiscal housekeeping.

The young king Ptolemy V is praised for cutting and cancelling taxes, forgiving arrears, and granting exemptions to the temples. In return, the priests agree to honour his cult. The stone belongs to a class of monuments, the donation stelae, that Egyptian rulers had been putting up for 2,000 years, each one essentially a publicly carved tax break for the priesthood. The key that unlocked an entire ancient language turned out to be a record of who got let off paying.

Rome, and the tax that funded a world

Rome ran the largest economy the western ancient world had built, and it taxed in exactly the places you would expect: at the border, on the land, and at the point of sale.

There was the portoria, a customs duty levied on goods crossing internal and external frontiers. There was the tributum, a levy on land and property in the provinces. And famously, there was tax farming. Rather than build a salaried collection service across a vast territory, the Republic auctioned the right to collect taxes in a region to private contractors called publicani. They paid the state an agreed sum up front and then kept whatever they could extract.

The arithmetic of this arrangement is grim and obvious. The contractor’s profit was the gap between what he collected and what he had paid, so his incentive ran entirely toward squeezing. The system raised money efficiently for Rome and made the tax collector one of the most hated figures in the ancient world, a reputation that survives in texts you may have read for entirely other reasons.

There was also the census, which we think of as a headcount and which was, in its origin, a tax instrument. You counted people and assessed their property precisely so you knew what each owed. To be counted was to be made liable. The word “census” carries the sense of assessment inside it.

What Rome could not do, despite all its sophistication, was tax income directly. There was no mechanism to measure what a Roman merchant or landowner actually earned in a year. So Rome taxed the harvest, the trade route, the sale, and the inheritance. It taxed the visible.

Salt, tithes, and the levies that built resentment

Two threads from the deep history run all the way forward into stories you already half-know.

The first is salt. For most of history, salt was the only reliable way to preserve food through a winter or a long voyage. Everyone needed it. It came from a small number of controllable sources, and demand barely moved with price. For a tax authority this is close to perfect. A salt tax is hard to avoid and easy to collect, which is exactly why rulers across China, Europe and India reached for it again and again over thousands of years. Keep salt in mind. It returns in Part 3 as the spark for one of the most famous acts of political defiance in modern history.

The second thread is the tithe, the customary payment of roughly a tenth of produce to a religious or feudal authority. Across medieval Europe the tithe funded the Church, and feudal levies funded the lord. The principle was the same as Uruk’s: a share of what the land produced, taken because the land’s produce could be seen and measured.

And running alongside every one of these systems was its shadow: resistance.

Tax has always been the point where ordinary people feel the state most directly in their lives, which makes it the place where they push back hardest. Tax revolts thread through the whole of recorded history, and they did something more than express anger. They began, slowly, to establish a principle. If a ruler needed money from his people, perhaps his people might ask for something in return. The idea that taxation and consent belong together was not handed down by philosophers. It was wrung out, levy by levy, by people who had had enough.

That idea is where Part 2 begins. For 5,000 years the state taxed what it could see and left what it could not. Then, inside a single lifetime, that changed, and the modern world of tax was born in the most expensive thing humans do. War.

Next, Part 2: how war taught the state to tax what you earn, and how a tax everyone hated became the thing that built modern political rights.

Series Guide - A History of Tax

• ✓ Part 1: What the First Writing Was For

• Part 2: The Tax That War Built

• Part 3: Kautilya to Gandhi

• Part 4: The Most Expensive 100 Rupees in History

  Subscribe now

  Share

  Leave a comment

Sources:

• Christopher Woods, “The Earliest Mesopotamian Writing,” in Visible Language: Inventions of Writing in the Ancient Middle East and Beyond (Oriental Institute, University of Chicago), on proto-cuneiform as administrative record and its origin in tokens and bullae

• The Origins of Writing, Department of Ancient Near Eastern Art, the Metropolitan Museum of Art, on the administrative function of early cuneiform

• British Museum and Britannica, on the content of the Rosetta Stone (Memphis Decree, 196 BCE), including its tax reductions, cancellation of arrears, and temple exemptions; on the donation-stelae tradition

• Kautilya, Arthashastra (R. Shamasastry translation, 1915), on the superintendent of commerce, customs and revenue (referenced fully in Part 3)

• Standard accounts of Roman taxation: portoria, tributum, the publicani, and the census as assessment instrument

  ---

Part 1 of this series was about a number the AI giants choose for themselves: how long they say their chips will last. Part 2A was about a number they cannot choose: the price their competitors charge. This part is about the money that keeps the whole thing moving while those two numbers argue with each other. The preceding parts are linked below:

Subscribe now

The industry is upside down.

Torsten Slok, chief economist at the investment firm Apollo, published a breakdown on 7 August 2026 that splits the AI industry into 4 layers and measures how much money each one makes. He worked from market data covering OpenAI, Anthropic, Microsoft, Amazon, Nvidia, AMD, Micron and others.

The chip and equipment makers earn a 41% profit margin. For every USD 100 of sales, USD 41 is profit. The layer that builds the models and sells AI to customers runs at negative 59%. For every USD 100 it takes in, it spends about USD 159.

In any normal business that is the wrong way round. The bakery should out-earn the flour mill. Here the flour mill is enormously profitable and the bakery cannot cover its costs.

Slok is blunt about what closes the gap. The AI boom’s profits, he wrote, are currently funded by investors rather than earned from customers.

Quick head’s up; the negative 59% is an estimate rather than a filed figure, since OpenAI and Anthropic publish no audited accounts, and it sits awkwardly beside a Reuters report that Anthropic expected its first quarterly operating profit this year. And it does not mean every sale loses money. It means the layer as a whole spends far more than it earns once the cost of building the models is counted alongside the cost of running them.

So: if the layer selling AI to customers cannot cover its costs, who keeps writing the cheques, and what do they get for it.

The one hard number

Almost every figure in this story is an estimate, a projection, or a run rate. One is not.

OpenAI’s audited 2025 accounts show revenue of USD 13.07 billion against total costs of USD 34 billion. That is a real filing, obtained by the writer Ed Zitron and verified independently by the Financial Times. We read that OpenAI went from USD 6 billion to USD 20 billion in a year, and Anthropic from about USD 1 billion to USD 9 billion. All 4 of those are annualized run rates: take one month’s revenue and multiply it by 12.

Run rate has a legitimate use. It tells you the pace a business has reached, which matters when a company is growing fast. What it is not is a year’s actual, audited earnings. During rapid growth it treats the best month as though the whole year looked like it, which is why the same year that produced a USD 20 billion run rate produced USD 13.07 billion of money actually received.

The growth is real. The size is smaller than the headline. And whichever number you use, this layer spends far more than it takes in, and investors make up the difference.

When your supplier is also your investor

There is a recurring arrangement everywhere in this industry.

• A company sells you equipment.

• Then it takes a stake in your business.

• Then, in some cases, it agrees to buy back whatever you cannot sell.

Each step is ordinary on its own. Stacked together they make a company’s sales look stronger and more independent than they are, because some of the demand is being created by the same people supplying the money.

The plainest example sits in Microsoft’s own annual report. Microsoft has put about USD 12 billion into OpenAI and owns roughly a quarter of it. In its accounts for the year to June 2026, Microsoft disclosed USD 24.1 billion of revenue from its commercial arrangements with OpenAI, including revenue-sharing payments. Separately, it disclosed that OpenAI owed it USD 6 billion on the day the year closed (30 June 2026).

Follow the loop. Microsoft puts money into OpenAI. OpenAI spends on Microsoft’s Azure cloud service. Microsoft records that spending as revenue, and it becomes part of the AI growth story Microsoft reports to its shareholders.

Why is this information public? So glad you asked. Accounting rules treat OpenAI as a related party, meaning a business Microsoft part-owns rather than an outside customer. Sales to an outside customer are just sales. Sales to a company you own a quarter of have to be flagged, because they do not prove the same thing about demand. The flour supplier again, at the scale of the largest software company on earth. Every business is owed money at any moment. But USD 6 billion against that year’s revenue is 91 days of sales, from one customer.

Then there is the behaviour. Microsoft had twice announced the size of its AI business as a headline: on pace for USD 13 billion in late 2024, USD 37 billion by March 2026. The quarter the related-party line appeared, it stopped announcing the figure.

Bloomberg calculates OpenAI accounted for around 70% of those AI sales. That is Bloomberg’s estimate, built on an assumption that growth rates held, not a Microsoft disclosure.

Nvidia and CoreWeave is the same arrangement, with the guarantee written down.

The supplier guarantees the demand

CoreWeave buys Nvidia chips and rents out the computing power by the hour.

On 9 September 2025 it filed a disclosure with the American market regulator describing an order from Nvidia worth USD 6.3 billion at the outset. Where CoreWeave’s capacity is not taken up by its own customers, Nvidia is obliged to buy the unsold capacity. The obligation runs to 13 April 2032. Nvidia held roughly 7% of CoreWeave’s shares as of mid-2025.

The flour supplier sold you the oven, took a stake in your shop, and promised to buy any bread you fail to sell. Your business now looks far safer to a lender. Whether anyone wants the bread is a separate question, and nothing here answers it.

That arrangement of USD 6.3 billion has since been proposed at 40 times the size.

OpenAI is planning a data centre campus in Pike County, Ohio. In late July, according to Wall Street Journal reporting, Nvidia was considering guaranteeing around USD 250 billion of its financing. This means standing behind the loans, so banks would lend to a company with no investment-grade credit rating. USD 250 billion is roughly a quarter of Nvidia’s stock market value.

Nvidia’s April 2026 filing discloses every guarantee of this kind it had ever given: maximum exposure across all of them, USD 3.5 billion. USD 3.5 billion across every partner. Four months later, USD 250 billion for one site.

Nvidia’s shares fell 5% when the terms were reported.

On 14 August the Journal reported the Nvidia cut the guarantee to under USD 120 billion, covering only the project’s 1st phase, after investors raised concerns about the risk. Neither company has commented and the reporting rests on unnamed sources, so treat the figures as reported rather than established.

A chipmaker was willing to put a quarter of its market value behind one customer’s borrowing. And investors stopped it within weeks. The brakes exist and were applied in public, this month.

Nvidia has walked back from roughly half the exposure it was considering. That risk did not disappear. Either someone else carries it, or the project gets smaller. The reporting does not say which.

The estimate is now being sold as a feature

On 10 August 2026, Nvidia announced financing partnerships with 6 of the largest investment firms in the world: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR.

The aim is to raise more than USD 500 billion of outside money for AI infrastructure, and to turn computing power (the machines that run AI, and the electricity to keep them running) into what the announcement calls an investable asset class, meaning something big institutions can buy and hold the way they buy property or toll roads. Nvidia describes the product to those investors as an investable asset, one which provides the lowest token cost, highest revenue and longest life.

Let’s explore that last piece: longest life. Part 1 argued that how long an AI chip lasts is an estimate, chosen by the companies whose profits it flatters, and that nobody outside the accounts department watches it. Nvidia is now selling that estimate as a feature to the people being asked to fund the buildout.

Depreciation is the cost of the machine, minus what you think it will fetch at the end, spread over its life. That middle number (cost minus what it will generate at the end of its life) is residual value. Part 1 argued it moves reported profit as effectively as useful life does, that no company discloses it separately, and that nobody is checking it.

Jensen Huang has now said Nvidia may provide residual-value support of up to 25% on some projects. In simple terms, Nvidia would promise lenders that if the chips backing a loan turn out to be worth far less than expected, it covers part of the shortfall.

That concedes 2 things.

1. The company that makes these chips does not know what they will be worth in 5 years either. If it did, it would not need to insure lenders against the answer.

2. And the number itself is now on the table. Residual value used to sit inside each company’s accounts, undisclosed and unquestioned. Nvidia has just attached a figure to it in public, in front of the institutions being asked to lend against these machines.

Ben Thompson, who writes the technology strategy publication Stratechery, calls the support “in a certain sense, a price cut”: Nvidia using its own profits to lower its customers’ cost of borrowing.

Remember these:

• The USD 500 billion is a set of memoranda of understanding, and Nvidia’s release says the partnerships still depend on final agreements. It is a headline resting on stated intentions, the same pattern as Anthropic’s USD 45 billion agreement with SpaceX, which either side can cancel with 90 days’ notice.

• And Apollo is on the list. Torsten Slok, whose margin analysis opens this article, is Apollo’s chief economist. His note was published on 7 August. Apollo signed on 10 August.

Both facts are public.

Whose money?

The answer is closer to home than the company names suggest. The chain runs like this.

1. A financing vehicle raises money and buys Nvidia chips and data centre equipment (computing power).

2. An AI company leases the computing power, creating a stream of payments.

3. The vehicle borrows against that stream.

4. Apollo, KKR and firms like them turn that borrowing into bonds and loans that other investors can buy.

These firms are no longer only buyout houses. Over the past decade they have bought or built insurance and annuity businesses, and they manage money for pension funds. So when they create this debt, they have somewhere ready to put it: the insurance reserves and retirement savings sitting on their own platforms.

Those pools are the target for a sound reason. Data centres are long-lived assets needing finance over many years, and insurers and pension funds have obligations stretching decades ahead. On paper it fits.

It is also a change in the kind of money at risk.

A year ago most of Big Tech was funding AI out of its own cash flows. Then came bonds: Goldman Sachs estimates AI-related financing now accounts for close to a quarter of all new investment-grade corporate borrowing in the United States, against AI investment approaching USD 600 billion this year. Now the search has reached insurance floats and pension funds.

Spending your own free cash flow is one thing. Tapping the debt markets is another. Bringing in money (insurance and pension funds) that exists specifically to be safe is a third.

Part 1 noted that anyone holding an international index fund already owns a piece of this. If, now, your pension or your insurer buys into these vehicles, you are funding the AI buildout whether or not you bought a share in it.

The strongest arguments against all of this

• Suppliers financing their customers is neither new nor sinister. Car makers have run lending arms for a century. Aircraft manufacturers help airlines finance planes.

• Nvidia guaranteeing CoreWeave’s unsold capacity lowers CoreWeave’s cost of borrowing and lets it build faster, which is an ordinary commercial objective. Microsoft’s USD 24.1 billion is public precisely because the accounting rules worked. A related-party transaction got flagged, as it is supposed to be. Nobody hid anything.

• Analysts make a third point aimed squarely at this article. Morgan Stanley’s Joseph Moore argues the financing platforms reduce concerns about money going in circles, because most of the capital comes from outside investors with Nvidia participating only to a limited extent.

• Bank of America’s Vivek Arya adds that these chips are unusually good collateral: they can be moved between operators if one fails, and Nvidia’s software keeps older chips useful for longer.

There is a problem. Vendor financing is ordinary when the customer’s business is proven and the financing smooths a purchase it would make anyway. It is different when the customer’s layer runs at a large loss and depends on repeated fundraising. Then the financing is not smoothing demand. It is substituting for it.

Disclosure surfaced the USD 24.1 billion (Microsoft’s June 2026 revenue from OpenAI), which is the system working. But a figure being disclosed tells you nothing about whether the arrangement behind it is durable.

The analysts’ case is the strongest, and outside capital does dilute the circularity. But note what Arya’s defence rests on: that the chips hold their value because they can be redeployed, and because software extends their useful life. That is an argument about residual value and useful life. The case for the financing being sound is built on the same 2 estimates that started this inquiry.

What the rest of the world is doing about it

One more source of capital holds this buildout up. Public money.

Governments watched American AI become expensive and Chinese AI become cheap and concluded that depending on either was uncomfortable. The answer everywhere carries the same name: sovereign AI, meaning artificial intelligence a country builds and controls itself rather than renting from someone else’s company.

• India’s national AI Mission runs on ₹10,372 crore over 5 years, roughly USD 1.2 billion. It has assembled 38,000 chips against an original target of 10,000, and it has produced real work. Sarvam AI, a Bengaluru startup the government chose to build India’s first homegrown model, trained entirely on Mission machines. BharatGen, a group of IITs and other institutes, launched a model that handles 22 Indian languages.

• Japan has committed about USD 6.2 billion over 5 years to a group of some 44 companies led by SoftBank, the Japanese technology investment group, alongside Sony, NEC and Honda.

• Europe is spending roughly EUR 20 billion of public money on large, shared data centres.

• France’s Mistral borrowed USD 830 million to build a data centre near Paris.

• Meta, one American company, plans to spend between USD 125 billion and USD 145 billion on capital projects this year. At the midpoint that is USD 370 million a day. Meta spends India’s entire 5-year national AI budget in a little over 3 days. That is not a criticism of the IndiaAI Mission, which beat its computing target nearly 4 times over under real constraints. It is a statement about scale.

And who’s selling the chips?

• India’s 38,000 chips are all imported, mostly from Nvidia.

• Japan’s programme announced the purchase of 27,500 of Nvidia’s newest chips, unveiled with Nvidia’s chief executive standing beside the minister.

• Mistral’s borrowed USD 830 million buys 13,800 Nvidia chips.

Outside China, which is barred from buying the best American chips and is building its own, every sovereign AI programme designed to reduce dependence on American technology begins by placing a large order with the same American company. On the national budget, it is independence. On Nvidia’s income statement, it is another customer.

That makes governments the third source of capital here, alongside the investors funding the model layer and the institutions Nvidia is assembling. Taxpayers are supporting the same buildout, under a different name.

What to watch

Part 2A listed the numbers that would tell you whether prices are collapsing. These are the ones that would tell you whether the financing is holding.

1. Related-party disclosures. Now that Microsoft has published its OpenAI revenue, the same line will appear every year. Watch whether that revenue grows faster or slower than the rest of the AI business, and whether the amount still unpaid at year end keeps rising.

2. Whether the memoranda become contracts. The USD 500 billion is currently a stated intention. Signed agreements, and money actually drawn, are a different thing.

3. What the guarantees cost, and who carries them. Nvidia’s obligation to buy CoreWeave’s unsold capacity is a real liability if the capacity goes unsold. On the Ohio project, watch for the final signed terms, and for who takes the portion of the financing risk Nvidia has stepped back from.

4. Whether new money keeps arriving. The layer losing money depends on repeated fundraising. A round that takes longer, prices lower, or attaches more conditions is the earliest signal available.

Epilogue:

The machines are real. The data centres exist, the chips are installed, the electricity is being drawn. None of that is in doubt.

What is open is the financing underneath. The layer selling AI to customers does not yet cover its costs, and the shortfall is met by investors. Some of those investors are the same companies selling the chips, guaranteeing the demand, and now assembling USD 500 billion of outside money to keep the cycle turning.

A supplier that finances its customer, guarantees that customer’s sales, and books the rising value of its stake in that customer as profit has built something that can look self-sustaining for as long as fresh capital keeps arriving.

Whether it is self-sustaining will be answered by whether ordinary paying customers grow into the space the investors are currently filling. I do not know how that resolves, and this is not a bubble call.

The infrastructure is the easy part to see. The financing is the omission.

Leave a comment

Share

Subscribe now

Sources

The one hard number and run rates: OpenAI’s audited 2025 accounts, obtained by Ed Zitron and verified independently by the Financial Times, for revenue of USD 13.07 billion against total costs of USD 34 billion; OpenAI full-year revenue of USD 3.7 billion (2024, CFO via Reuters); OpenAI CFO Sarah Friar’s January 2026 statement of USD 6 billion and USD 20 billion annualized run rates.

Margins: Torsten Slok, Apollo Global Management, “The AI value chain,” 7 August 2026, for the 41% profit margin at the silicon and equipment layer and the minus 59% operating margin at the models and applications layer, drawn from Pitchbook and Bloomberg data, and for the observation that upstream margins are funded by capital raised downstream; reported by Fortune, 10 August 2026. Reuters (May 2026) on Anthropic’s expected first quarterly operating profit.

Microsoft and OpenAI: Microsoft Corporation Form 10-K for the fiscal year ended 30 June 2026, related-party disclosure under ASC 850, for the USD 24.1 billion of revenue from commercial arrangements with OpenAI inclusive of revenue-sharing payments, the USD 6.0 billion of accounts receivable at 30 June 2026, the approximate 25% as-converted interest disclosed in that filing, and total funding commitments of USD 13 billion of which USD 11.9 billion had been funded. Microsoft’s stake was disclosed as approximately 27% in its 10-Q filings of October 2025 and March 2026, and was diluted by OpenAI’s March 2026 funding round; the 25% figure is the one in the same annual report as the USD 24.1 billion. Bloomberg, “Microsoft’s AI Sales Mostly Come From OpenAI, Disclosures Show” (August 2026), for the estimate that OpenAI accounted for about 70% of Microsoft’s fiscal 2026 AI sales, resting on Bloomberg’s assumption that the 123% run-rate growth reported for March continued, and for Microsoft’s USD 13 billion and USD 37 billion prior AI run-rate disclosures and the absence of an update at the fourth quarter.

Nvidia’s CoreWeave backstop: CoreWeave, Inc. Form 8-K filed 9 September 2025, disclosing an order form under the Master Services Agreement dated 10 April 2023, initial value USD 6.3 billion, under which Nvidia is obligated to purchase residual unsold capacity through 13 April 2032; reported by Reuters and CNBC, 15 September 2025. Nvidia’s approximately 7% holding of CoreWeave Class A shares as at 30 June 2025, per CNBC; that holding may have changed since.

Nvidia’s existing guarantee exposure: NVIDIA Corporation Form 10-Q for the quarter ended 26 April 2026, Note 8, disclosing agreements to guarantee partners’ facility lease obligations in exchange for warrants, with maximum gross exposure under all such agreements of USD 3.5 billion, reduced as partners make payments over terms of 5 to 7 years, and USD 712 million placed in escrow by those partners.

The Ohio guarantee: The Wall Street Journal, 14 August 2026, citing people familiar with the discussions, for the reduction of Nvidia’s proposed financing guarantee for OpenAI’s Pike County, Ohio campus from approximately USD 250 billion to under USD 120 billion, the restriction of the backstop to the project’s first phase, and investor concern as the stated reason; reported by Reuters, 14 August 2026. Nvidia’s 5% share price fall when the original terms were reported, and the observation that USD 250 billion approximated a quarter of Nvidia’s market capitalisation, from the same reporting. Neither Nvidia nor OpenAI has commented publicly, and no filing on these terms had appeared at the time of writing.

Residual-value support and the capital chain: Fortune, Eva Roytburg, “Nvidia found a new way to keep the AI boom funded: your retirement money,” 12 August 2026, for Jensen Huang’s statement that Nvidia may provide residual-value support of up to 25% on some projects, the description of the financing chain from vehicle to insurance and retirement capital, Goldman Sachs’s estimate that AI-related financing accounts for close to a quarter of gross US investment-grade issuance against AI investment approaching USD 600 billion this year, Ben Thompson’s characterisation of the support as a price cut and his concern about safety-seeking capital, Morgan Stanley’s Joseph Moore on third-party capital reducing circularity, and Bank of America’s Vivek Arya on GPU redeployability and software-extended useful life.

Compute as an asset class: NVIDIA press release, “NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over USD 500 Billion of Third-Party Capital,” 10 August 2026, for the partner list, the investable asset class framing, the “lowest token cost, highest revenue and longest life” description, and the statement that the partnerships remain subject to execution of final agreements; BBC News and CNBC coverage, 10 August 2026; InfotechLead, 11 August 2026, for Goldman Sachs’s interest in a credit market backed by Nvidia compute.

Sovereign AI: IndiaAI Mission budget and GPU figures from MeitY statements and Lok Sabha replies (2026); Sarvam AI and BharatGen model launches at the India AI Impact Summit (February 2026); Japan’s METI commitment of ¥1 trillion and the Rubin GPU procurement announcement (July 2026); EU AI gigafactories programme (~EUR 20 billion); Mistral AI’s USD 830 million debt financing, as reported March 2026; Meta Q1 2026 results for 2026 capital expenditure guidance. US export controls on advanced AI chips to China, in force since 2022 and extended since.

2 things changed under salaried professionals in 2026, and most people have noticed neither.

Subscribe now

• The first is that the Income Tax Act, 1961 is gone. The Income Tax Act, 2025 replaced it entirely with effect from 1 April 2026, along with new Income Tax Rules, 2026. Rates, limits and benefits carried over almost untouched. The section numbers did not. Section 80C is now Section 123 read with Schedule XV. Section 80CCD is Section 124. Section 80D is Section 126. “Previous Year” and “Assessment Year” have been replaced by a single “Tax Year”. Form 12BB, the investment declaration you submit to your employer, is now Form 124. None of this changes what you owe. All of it changes what you type on a form, what your payroll system needs to have been updated to, and what a notice will reference when one arrives.

• The second change is the one that costs money if you miss it. The new regime is the default. You are on it unless you actively opt out.

Which regime, and why the answer probably moved:

At your previous salary one of these was clearly better. A promotion is one of the few events that can genuinely flip the answer, because the old regime only wins when your deductions are large enough to overcome the new regime’s lower rates, and both sides of that comparison move when your income moves.

Run both. Not an estimate. The actual calculation, with your actual rent, your actual PF, your actual insurance premiums.

The one deduction that survives the new regime

Almost every deduction you have heard of belongs to the old regime and disappears if you stay on the default. Section 123 (formerly 80C). Section 126 (formerly 80D). HRA. One survives, and it is the most useful lever available to a promoted employee, and hardly anybody uses it.

Section 124(2), formerly Section 80CCD(2), allows a deduction for your employer’s contribution to your NPS (National Pension System) Tier-I account, up to 14% of your basic plus dearness allowance, and it is available under both regimes.

This is your employer routing part of your CTC into your retirement account instead of paying it to you as taxable salary. It reduces your taxable income rupee for rupee.

On a basic of 7.5 lakh, the full 14% is INR 1,05,000 of income made tax-free. Used deliberately, it can pull taxable income back below the INR 12 lakh line and eliminate your tax bill entirely.

2 conditions:

• Your employer has to offer corporate NPS and agree to restructure, which many will do on request and almost none will offer unprompted AND

• Employer contributions to NPS, EPF and superannuation combined must stay under INR 7.5 lakh a year, or the excess becomes a taxable perquisite.

If the old regime still wins for you

Then the deduction cupboard is worth opening properly, and it is larger than most people work it.

1. Section 123 allows INR 1.5 lakh, but your EPF contribution already sits inside that limit. Subtract it first and fill only the genuine gap. ELSS (equity linked savings schemes, tax-saving mutual funds with a 3-year lock-in) is the most capital-efficient way to do that.

2. Section 126 covers health insurance for you, your spouse, your children and your parents, with an additional INR 50,000 available if your parents are senior citizens.

3. Section 124(1B), formerly 80CCD(1B), gives you INR 50,000 for your own NPS contribution over and above the INR 1.5 lakh cap. At the old regime’s 30% slab with cess, that single deduction saves INR 15,600.

Stack all 3 at their maximum and the saving runs from INR 46,800 at the 20% slab to INR 85,800 at the 30% slab with senior citizen parents covered. That is a genuine range, computed at INR 1.5 lakh plus 50,000 plus INR 25,000, and INR 75,000 of Section 126 where parents qualify.

HRA just changed, and it may have changed for you

For more than 40 years, only Mumbai, Delhi, Kolkata and Chennai qualified for the higher 50% HRA limit. Under Rule 279 of the Income Tax Rules, 2026, Bengaluru, Hyderabad, Pune and Ahmedabad joined them with effect from Tax Year 2026-27.

If you live in one of those cities, pay rent, and are on the old regime, your exemption ceiling moved from 40% of salary to 50%. Whether that translates into actual saving depends on which of the 3 limbs binds, because the exemption is the lowest of: HRA actually received, rent paid minus 10% of salary, and 50% or 40% of salary. If the second limb is your constraint, the city change gives you nothing.

Two documentation points that catch people:

1. Landlord PAN is mandatory once annual rent crosses INR 1 lakh

2. Form 124 now requires you to disclose your relationship with your landlord, which matters a great deal if you pay rent to your parents. That arrangement is entirely legitimate when they own the property, but: they must declare the rent as income on their own return. Pay by bank transfer. Keep receipts monthly.

The March problem

The annual scramble of poorly chosen tax-saving purchases made under deadline pressure happens for a single reason: the planning was postponed until the deadline made it urgent. The fix costs one hour in April.

If your income has crossed INR 15 lakh, one session with a fee-only financial adviser is worth the outlay. Fee-only means they charge for advice and earn nothing from what they recommend. Expect INR 3,000 to INR 5,000 for a review of your tax position, insurance adequacy and asset allocation. The most expensive financial advice is the free kind, and the invoice arrives in instalments over 20 years.

What to do right now:

1. Compute your liability under both regimes at your promoted salary, using your real deductions. Choose on the number, not on habit.

2. Ask HR whether corporate NPS is available, and whether up to 14% of basic can be routed into it under Section 124(2). This works on both regimes and is the single highest-leverage question in this article.

3. Submit Form 124, the replacement for Form 12BB, within the first week of your promoted salary. An outdated declaration means excess TDS from month one and a refund you wait a year for.

4. If you are staying on the old regime: subtract your annual EPF from 1,50,000 and fill only the remainder. Claim Section 126 for every premium including your parents’. Add the 50,000 NPS deduction under Section 124(1B).

5. If you live in Bengaluru, Hyderabad, Pune or Ahmedabad and are on the old regime, rework your HRA at the 50% limit and check whether it changes your regime choice.

6. Fix your rent documentation now: agreement, monthly receipts, bank transfers, landlord PAN above 1 lakh annual rent, and the relationship disclosure on Form 124.

Subscribe now

Share

Leave a comment

Part 1 of this series looked at how the AI giants report their profits and found the most important number in the accounts is one they choose for themselves: how long they say their chips will last. Spread the cost of a chip over 6 years instead of 3 and the profits look healthier, and the estimate is made by the very people those profits flatter.

You do not need to have read it. Here is all it assumed, in one line. A reported profit is an estimate built on other estimates, and an estimate can be wrong.

This part is about a number the AI giants cannot choose for themselves. The price their competitors charge.

Subscribe now

Start with what AI actually costs to buy. It is sold by the word, more or less. When a company plugs AI into its app, it pays the provider a small amount for every chunk of text going in and coming out. The industry calls these chunks tokens, each roughly three quarters of a word, and charges by the million. You pay per unit, the way you pay for electricity.

2 price tags from July 2026, both from the providers’ own published price lists.

• OpenAI (the maker of ChatGPT) charges USD 5.00 per million tokens of input for its flagship model, GPT-5.5.

• DeepSeek (a Chinese AI company) charges USD 0.14 for its V4 Flash model. The same unit. About 35 times cheaper.

Both figures are input rates, meaning what you pay to send text to the model. Output costs more, and bulk or repeat traffic costs less. These prices move monthly, so check the providers’ own pages on the day you read this.

DeepSeek’s model is not as capable as OpenAI’s best. It does not need to be.

Most of what companies actually use AI for is summarising documents, answering customer questions, sorting and labelling text. That work needs a model that is good enough, at a price that does not wreck the budget.

Part 1 argued that the AI boom’s reported profits depend on a number the companies choose themselves: how long they say their chips will last. This part is about a number they cannot choose. The price a competitor charges.

Nearly a 1/3 of the work, 1/25 of the money

Vercel runs the websites and apps of thousands of companies, which means it can see which AI models those apps actually call. Its June 2026 data contains the following:

• Open-weight models, meaning models whose trained recipe is published so anyone can download and run them on their own machines, handled 29% of all AI traffic on the platform.

• They collected under 4% of the money.

• Most of the leading open-weight models are Chinese.

For customers, the cheap models genuinely are that cheap. Companies route enormous volumes of everyday work through them and barely pay anything.

For the American AI companies, it is worse than it looks. Revenue is holding because they still own the expensive, high-stakes work. The volume that could have become tomorrow’s revenue is moving elsewhere.

OpenRouter is a marketplace where software developers pick between AI models.

• American models fell from roughly 70% of its traffic to roughly 30% in the 12 months to June 2026, per a Bloomberg chart built on its data.

• A CNBC investigation in early July found Chinese models had carried at least 30% of American companies’ traffic there every week since February, peaking at 46%, against a prior 12-month average of 11%. By late July the reported figure had reached 57% in a single week.

Both are developer platforms, and that matters. The people using them (tracked by Vercel and OpenRouter) are price-sensitive and free to switch on at any time, which large companies bound by contracts and compliance rules are not. Read those figures as a leading indicator of where usage patterns are heading, not as enterprise market share today.

The sturdier number comes from IDC, a technology research firm, which surveyed 260 decision-makers at American companies employing more than 1,000 people. 47% said they used a Chinese-made model for at least one purpose. 1 in 5 reported extensive use.

Why customers are switching

It’s really just the money.

American AI companies have been raising prices while their Chinese competitors cut them.

One pricing tracker that follows both puts DeepSeek’s rates down roughly 75% over the past year, while OpenAI’s flagship input price rose 4-fold. Ramp processes corporate expenses, so it can see what businesses actually pay for, and it reported DeepSeek becoming the top trending software vendor on its index. The pressure from the buyers switching to cheaper models is now measured rather than anecdotal.

KPMG surveyed 2,145 senior leaders across 20 countries in the second quarter of 2026, all at organisations earning more than USD 50 million a year.

• 49% said they had narrowed, delayed or paused their AI agent deployments because the running costs had outgrown the value produced.

• Only 7% of those leaders said their organisation had reached an established return on its AI investment.

• Planned AI investment held at an average of USD 188 million over the following 12 months, essentially flat against the previous quarter.

• The share of companies running several agents together doubled from 9% to 18%.

• AI remained a top investment priority for 79% of leaders, up from 74%.

Adoption is accelerating and proof of return is not arriving. It is also a survey of what executives say rather than an audit of what happened, Opinions and facts may diverge.

Most buyers cannot see what they are spending while they spend it.

In KPMG’s companion American survey, of 204 leaders at companies above USD 1 billion in revenue, only 26% had real-time visibility into what their AI costs to run, and only 36% had direct controls on token usage. They are buying a metered utility without a meter.

The expensive models are not being abandoned. They are being demoted.

Wall Street Journal reporting in late July described how companies now assemble several models into one workflow. A cheap model does the bulk of the work. An expensive one plans it or checks it at the end.

Telnyx, which builds software for automated assistants, runs 1,400 of them with this workflow:

• A cheap Chinese model doing the bulk of the work

• An Anthropic model directing it

• An OpenAI model checking the result.

Volumes are handled by the cheaper models. Quality control is still managed by more expensive ones.

Another interesting wrinkle. Both American labs are increasingly using free usage and credits to keep customers from leaving. Marty Kausas, chief executive of the customer support company Pylon, told the Journal there is no loyalty left in the market and called the current period a bloodbath. His company has been handed months of unlimited free usage, including roughly USD 1.6 million of free tokens from a single vendor this year.

A young company gives its product away to win customers. An established one gives it away to stop them leaving, and this current pattern is the second kind. It also does something specific to the accounts. Revenue can hold steady while the quality of that revenue deteriorates underneath it, because the same reported figure now leans on discounts, credits and giveaways that were not needed a year ago. The American AI companies are less than enthusiastic about this trend.

On 24 July 2026, 25 companies signed an open letter, “Open Weights and American AI Leadership”, urging Washington not to restrict open-weight models. The signatories included Nvidia, Microsoft, Meta, Mistral, IBM and the venture firm Andreessen Horowitz. Nvidia’s chief executive fronted it with the first post of his life on X. 3 names declined to sign. OpenAI, Anthropic and Google, the 3 leaders in expensive closed models.

The split makes economic sense. Open models run on Nvidia’s chips whoever builds them, so the chipmaker gains as they spread. Expensive intelligence is the product OpenAI and Anthropic sell, so they gain as it stays scarce.

There is no subtlety in play here.

Why this breaks into Part 1’s arithmetic

Part 1’s dial: the AI giants spread the cost of their chips over 5 to 6 years, and their reported profits depend on that choice. A 6-year write-off does not just assume the chip survives 6 years. It assumes the chip can earn for 6 years. That is the price problem.

Those are different claims. Back to the bakery. You bought the oven expecting to sell bread at INR 50 a loaf for 6 years. A shop opens down the road selling at INR 1.50. Your oven still works perfectly. Your arithmetic does not.

Bloomberg reported last month that the industry’s AI revenue finally edged past its depreciation charge, USD 25 billion against USD 21 billion in a quarter. That milestone rests on both numbers behaving. The depreciation figure depends on the useful-life estimate, which Part 1 showed is contested. The revenue figure depends on prices holding, and a price war is attacking it now.

The companies spending hundreds of billions on infrastructure are making a bet that AI services will generate enough revenue to justify that spending. The people running the actual traffic are voting, 29% of the volume at a time, that it will not.

The thing to be wary about:

The platforms measuring the shift are developer marketplaces, where users are price-sensitive and free to switch. The corporate world is stickier. Ramp’s own economists note that Chinese model adoption across American businesses overall remains low, concentrated in a handful of AI-heavy firms.

Banks, hospitals, defence contractors and governments face much higher barriers to sending sensitive data to a model hosted in China.

That is less protection than it sounds. Because the trained model itself is published, a European bank can download Alibaba’s Qwen models and run them on its own servers in Frankfurt, and a UBS report describes a large global bank doing exactly that. The data never leaves the building. The revenue never reaches the American providers. A Chinese model does not require a Chinese data centre.

Even buying one is difficult. Most big companies get their AI through 3 online stores, run by Amazon, Microsoft and Google, and none of them carries these Chinese models yet. To use one, a company has to set it up on its own computers instead.

Politics makes it harder still. A committee in the US Congress is looking into American companies that use Chinese AI, and the research firm Gartner has warned clients that admitting to it in public can cost them business. For many firms, the savings are not worth the bother.

Quality still commands money. Anthropic (the maker of Claude, OpenAI’s closest American rival) reportedly told investors it expected its first quarterly operating profit this year, which is not what a company being destroyed by cheap competition looks like.

And a price that fell 75% in a year can keep falling, including below its own cost. Nobody outside China knows what these models cost to run, because these companies do not publish audited accounts the way listed American firms must. Whether the cuts reflect genuine efficiency, thin margins, state subsidy or a blend is not visible from here, and a price war fought with subsidised prices can reverse.

The argument in this piece does not need Chinese prices to stay low forever. It needs American pricing power to hold for 6 years, which is a stronger assumption than it looked a year ago.

The best argument against all of this

When something becomes much cheaper, people usually use much more of it.

The Victorian economist William Stanley Jevons noticed in 1865 that more efficient steam engines increased Britain’s coal consumption instead of reducing it, and the pattern has recurred often enough to carry his name. Cheaper bread does not mean the bakery earns less. It usually means the bakery sells a great deal more bread.

Applied here, Jevons’ paradox says the price collapse is the whole point. Cheaper output means more uses become affordable, volume explodes, and total revenue grows even as the price per unit falls.

Satya Nadella, Microsoft’s chief executive, reached for exactly this argument within hours of the DeepSeek crash of January 2025, when Nvidia lost USD 589 billion of market value in a single day. That remains the largest one-day fall for any company in American history.

The evidence supports him.

The Silicon Valley venture firm Andreessen Horowitz tracks what it calls LLMflation: for a fixed level of capability, prices have fallen roughly 10-fold a year since 2021. A model of the quality of GPT-3 cost USD 60 per million tokens in 2021 and USD 0.06 by late 2024. Across the same period, revenue at the American AI companies multiplied several times over.

Cheaper output did not shrink the market. It enlarged it.

People cite OpenAI going from USD 6 billion to USD 20 billion, and Anthropic from about USD 1 billion to USD 9 billion. All 4 of those are annualized run rates, which take a single month’s revenue and multiply it by 12. During fast growth that is the most flattering way to state revenue, because it treats the best month as though the whole year looked like it.

OpenAI’s real full-year revenue was USD 3.7 billion in 2024 and USD 13.07 billion in 2025, the second figure from audited accounts. The growth is real. The magnitude of the growth is not in keeping with the expected numbers.

That argument is correct about volume and but silent about capital.

Take 2025 on its own.

Prices for a given capability fell by roughly 10 times on Andreessen Horowitz’s measure. OpenAI’s audited revenue rose from USD 3.7 billion to USD 13.07 billion, about 3.5 times. For revenue to rise 3.5 times while the price of a unit falls 10 times, the number of units sold has to rise around 35 times. Every one of those extra units needs a chip, a building, power and cooling.

Revenue grew 3.5 times. The equipment needed to produce it grew far faster than that.

Depreciation is not an aggregate question. It is asked of each machine separately. Will this particular asset earn back what it cost before it is written off? A market can triple in size while every chip inside it fails that test. The bakery can sell 35 times more bread and still not cover the oven, if the price per loaf fell further than the extra sales made up for.

So Jevons’ paradox is a sound reason to expect the AI market to keep growing. It is no reason to expect a given chip to earn back its cost over 6 years. Those 2 claims are constantly treated as one.

That gap, between a growing market and a chip that pays for itself, is the whole price problem. And it is only half the story.

The assumption underneath the whole buildout is this: that a chip bought today will keep earning at today’s prices for the 5 or 6 years it is written off over. The price war is the evidence that it will not. Prices are already falling faster than the chips are ageing.

So, if the machines will earn less than the plan assumes, the next question is who keeps writing the cheques in the meantime. That is Part 2B.

Share

Subscribe now

Leave a comment

Sources

Pricing and adoption: official provider pricing pages for OpenAI and DeepSeek (as of July 2026, and subject to monthly change); Vercel AI Gateway Production Index (June 2026) on open-weight token and spend share; a Bloomberg chart built on OpenRouter data for US model token share; CNBC, “Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge” (7 July 2026), for the weekly US enterprise token-share figures; IDC survey of 260 US decision-makers at companies with more than 1,000 employees, as reported by The Daily Upside, “Modelmaxxing Replaces Tokenmaxxing” (25 July 2026); one pricing tracker following both providers for the DeepSeek 75% decline and the OpenAI 4-fold rise, attributed rather than stated flat; Ramp trending-vendor index and Ramp economist commentary on overall US adoption.

Demand side: KPMG Global AI Pulse Q2 2026, published 24 June 2026, surveying 2,145 senior leaders in 20 countries at organisations above USD 50 million revenue, for the 49% pullback, the 7% reporting established return, the USD 188 million planned investment, the 79% priority figure and the 9% to 18% multi-agent shift; KPMG’s companion US survey of 204 leaders at companies above USD 1 billion, for the 26% cost-visibility and 36% token-control figures; reported by Reuters Breakingviews (7 August 2026), CFO Dive and Forbes. GitHub Copilot’s move to usage-based billing and the projected USD 180 monthly bill: Visual Studio Magazine, cited by Forbes, August 2026, offered as one writer’s projection from a single day rather than as evidence.

Model mixing: The Wall Street Journal, “How US companies flipped from tokenmaxxing to thrift-maxxing” (26 July 2026), for the Telnyx multi-model workflow, reported behind a paywall and not independently confirmed.

Distribution and friction: absence of Chinese open-weight models from Amazon, Microsoft and Google’s AI marketplaces, per Bloomberg Intelligence analysts Mandeep Singh and William Tong (July 2026); CNBC, “Lawmakers probe growing use of Chinese AI models in U.S. companies” (8 July 2026); Gartner client guidance on procurement resistance, as reported July 2026.

The counter-argument: Guido Appenzeller, “Welcome to LLMflation,” Andreessen Horowitz (November 2024), for the 10-fold annual price decline and the USD 60 to USD 0.06 figures; CNBC and Forbes on Nvidia’s USD 589 billion single-day loss (27 January 2025); Satya Nadella’s public invocation of Jevons’ paradox (January 2025); OpenAI full-year revenue of USD 3.7 billion (2024, CFO via Reuters) and USD 13.07 billion (2025, audited accounts via Ed Zitron and the Financial Times); OpenAI CFO Sarah Friar’s January 2026 statement of USD 6 billion and USD 20 billion annualized run rates.

Depreciation link: Bloomberg on AI revenue exceeding depreciation, drawing on Exponential View (June 2026); see Part 1 for the useful-life analysis.

I have watched a lot of people lose their tempers over the years.

In classrooms, in offices, across dinner tables, in traffic, in meetings where the stakes were high and the patience was thin. I have watched it up close and from a distance, in people I respected and people I didn’t, in situations that were genuinely terrible and situations that were merely inconvenient.

And I have not once, in all of that watching, seen a tantrum solve a problem.

Not one. Not partially. Not even accidentally.

What I have seen is the aftermath. The energy spent. The relationships strained. The original problem still sitting exactly where it was before the losing of the temper, now with the additional company of whatever damage the outburst left behind. The problem didn’t move. The person just exhausted themselves failing to move it.

This is how I arrived at a conclusion that has shaped everything since. We have limited energy. A finite, non-negotiable daily allocation that does not expand because we need more of it. The question is never whether we have enough energy. The question is always where it goes.

I can use it to do the thing. Or I can use it to feel things loudly about the thing. But the thing, in both cases, remains. And in one of those cases it also gets done.

The internal experience of Calm Intensity is difficult to describe to someone who hasn’t felt it. The closest I can get is this.

Imagine switching on a spotlight in a dark room. Everything the beam touches becomes completely clear, completely present, completely the only thing that exists. Everything outside the beam doesn’t disappear. It simply fades. It will still be there when you move the light. But right now, it has no claim on your attention. That’s it. That’s the whole thing.

Not a floodlight filling the room with diffuse brightness. Not ambient light that illuminates everything equally and therefore nothing particularly. A spotlight. Deliberate, directional, chosen.

My brain, when it engages this way, runs in a rhythm I can only describe as staccato. Identify. Complete. Next. Rest. Repeat. Again. There is no commentary between the steps. No internal performance review mid-task. No feelings about the feelings. Just the clean sequence, running.

It is not a dramatic state. It doesn’t feel like intensity from the inside, which is perhaps the most important thing to understand about it. From the outside it may look like focus. From the inside it feels like the opposite of effort. Like everything unnecessary has simply been set down and what remains is just the work and the doing of it.

I want to be honest about something, because without this honesty the whole thing sounds like a manual written by someone who has never had a boring day.

Sometimes the spotlight switches on and illuminates something that is boring AF.

A form that needs filing. A follow-up that needs sending. The same chapter read for the third time because it still needs to go in. Day 187 of an 8-month medication protocol when the novelty of surviving a TB diagnosis has thoroughly worn off and it is just another set of pills on another unremarkable morning.

Calm Intensity is not reserved for the dramatic or the interesting or the things that deserve enthusiasm. The staccato doesn’t have a setting for “but this is dull.” It just runs. Identify. Complete. Next. Rest. Repeat. Again.

This is the actual muscle. Not the ability to focus intensely on something genuinely engaging. Anyone can do that. That’s just interest, not discipline.

The real training is switching the spotlight on for the thing that doesn’t deserve it and doing it anyway, with the same quality of attention you’d give something that mattered enormously, because at some level everything on the list made the list for a reason.

And here is what I have learned about the energy equation after years of running it.

It costs nothing to focus. Nothing. The spotlight uses no more energy than the floodlight. Less, arguably, because it isn’t trying to illuminate everything simultaneously.

What is expensive is the frittering. The half-attention that produces half-results. The task switching that burns energy between tracks without moving either forward. The internal negotiation about whether to begin, the renegotiation about whether to continue, the running commentary about how long it’s taking and whether it’s working and whether there might be a better approach.

All of that costs something real. And the thing still isn’t done.

Maximum effort is not the same as effective effort. Intensity without direction is just noise and heat. Control without intensity doesn’t move anything. But when the two meet, when you bring the full beam of your attention to the one thing that needs it, something shifts.

The work doesn’t become easier or more interesting or more rewarding in any obvious way. It just gets done. And then the next thing gets done. And somewhere in that accumulation, glacially, inevitably, things start to move.

I am human. The spotlight doesn’t always switch on smoothly. Some days the boring AF thing sits there longer than it should, and the staccato takes longer to engage. Some days the energy allocation feels smaller than the list requires.

On those days I do not perform intensity I don’t have. I find the one thing on the list that will move the needle and I give it whatever the spotlight has available. Not perfect focus. Just honest focus. Directed at the right thing.

That has always been enough. The tantrum, for the record, is still zero for everything.

Between 1995 and 2000, the internet arrived and investors lost their heads. Money poured into any company with a .com in its name, most of which had no profits and a few of which had no product. The Nasdaq, America’s technology stock index, rose more than 6-fold in 5 years, peaked in March 2000, and then fell 78% over the next 2 and a half years.

Thousands of companies vanished. This was the dot-com bubble.

The internet itself was entirely real. But being right about the technology and being right about the price turned out to be 2 different things.

That word is back. Barely a week passes without someone declaring the AI economy a bubble, and someone equally certain declaring it is not. Both camps are arguing about valuations, which is a question about the future, and no one outside these companies can settle it.

So I went looking for a smaller question with a checkable answer. Valuations are a guess. Reported profits are a matter of record, filed and audited. I wanted to know whether that record says what it appears to say. It starts with a decision 2 companies made in the same month, in opposite directions.

Subscribe now

The first 2 estimates - Depreciation & Useful Life:

In January 2025, Amazon and Meta each decided how long their servers would last. Amazon said 5 years. Meta said 5.5. Amazon charged itself USD 1.3 billion. Meta credited itself USD 2.9 billion. Both were permitted, and both were audited. Same month, same machines, same supplier but diverging decisions.

Imagine for a moment, that you are a baker and you buy an over for your bakery. It costs INR 6 lakh. You do not treat that as an INR 6 lakh loss in the month you buy it, because you will use the oven for years. So you spread the cost. If you expect 6 years of use, you charge INR 1 lakh a year against your profits. The value of the oven in your books falls by INR 1 lakh every year. This annual, pro-rated hit to your profits is called depreciation.

Now suppose the oven goes belly-up after 3 years. Nothing else changed in your kitchen. You realise that your books, though accurate up to that point, based on assumptions you had made, had been telling you a falsely comforting story. Your accounts say the oven is still worth INR 3 lakh after 3 years. You know it is worth nothing, so you take that INR 3 lakh hit in the current year. For 3 years you reported profits better than the truth, and now you need a new oven you had not budgeted for.

How long you say the machine lasts is called its useful life. It is an estimate. Accounting rules give you benchmarks, but you make the final call on your own judgement, and your auditor tests whether it is reasonable.

Here is why it is this concept, applied to the AI industry, becomes your problem. The 7 largest technology companies make up roughly 38% of the Nasdaq 100, the index of the largest companies on America’s technology-heavy exchange. If you hold an international index fund, close to 40% of it is 7 companies whose reported profits depend on an estimate you have never been shown.

I am going to keep coming back to the bakery, because every move in this story has an oven-sized version.

The estimate driving profits:

Last month Bloomberg reported a milestone, drawn from research by Exponential View (a technology research firm).

Across the big cloud companies, global AI sales outside China finally rose above what those same companies charged as depreciation on their chips and data centres. It was the second quarter running. In bakery terms: the bread is finally paying for the oven.

Half of that comparison is a sales figure, which is countable. The other half is the estimate (depreciation).

Big cloud companies write off AI hardware over roughly 5 to 6 years. Nvidia (the American company making most of the world’s AI chips) now ships a new generation every single year.

Michael Burry (the investor who called the 2008 American housing collapse) argued in November 2025 that the real working life is closer to 2 or 3 years. Jim Chanos (a short seller known for exposing Enron) and Aswath Damodaran (a valuation professor at New York University) have made versions of the same argument.

On Burry’s figures, Oracle’s and Meta’s operating profits would be overstated by more than 20% by 2028.

Move the dial from 6 years to 3 and depreciation roughly doubles. Revenue no longer outpaces depreciation.

The reality is different from this inital mismatch. According to Nvidia, customers see 4 to 6 years in practice. Chips bought in 2020 are still running commercial work today and still earning money. They earn less each year as newer chips arrive. Your 5-year-old oven still bakes; it just cannot keep up with the shop that bought the new one.

An asset that keeps working while its earning power falls is the textbook case for depreciating it. The only question is whether 6 equal annual charges match a curve that drops faster at the start. On the evidence so far, they do not.

The second estimate - Useful Life:

There is a second estimate underneath, and nobody discloses it. Back to the bakery. Before you can spread that INR 6 lakh, you have to settle a second question: what the old oven will fetch when you are done with it.

If you think it is scrap, you spread the full INR 6 lakh over 6 years. If you think a smaller bakery will pay INR 1 lakh for it, you only spread INR 5 lakh, and your annual charge drops. Same oven, same 6 years, smaller hit to your profits. That second number of INR 1 lakh, is called residual value.

Annual depreciation is cost, minus residual value, divided by useful life. Two inputs, not one.

Every argument in public is about the years. Burry, Chanos, Damodaran, the analysts, the newsletters, all of them are arguing about the denominator. The other input sits there untouched. It should not.

Servers were historically assumed to be worth close to nothing at the end, because a 6-year-old machine was scrap. That assumption is now under pressure from an unexpected direction.

Computing power has developed a scarcity price and the beginnings of a resale market, and Meta is reported to be building a business to sell its own spare capacity. That resale market is genuinely good news for these companies. It is the strongest argument that this build is unlike the dot-com era, when the fibre optic cable laid in the boom sat unused for years afterwards. Data centres are starting to behave like assets you can trade rather than money you have sunk.

It also hands every finance department in the sector a documented reason to say that their used chips are worth something. Accounting standards require residual value to be reviewed every year and adjusted when market conditions change. A company revising it upward is doing exactly what the rules ask. The difference is who notices.

A change in useful life gets announced. Amazon announced the change from 6 to 5 years on an earnings call. Meta put 5.5 years in a press release. Analysts had it modelled within hours. A change in what a company assumes its old servers will fetch, the residual value, is not separately disclosed by any of these companies, and nobody is asking. Same effect on profit. None of the scrutiny.

I am not alleging anyone has done this. The lever exists, it is invisible from outside, and it is the one nobody is checking.

The people who set the number are very the people it flatters:

Let’s return to January 2025. Amazon shortened the life of some servers and networking equipment from 6 years to 5 and said why in its own filing: the increased pace of technology development, particularly in artificial intelligence.

Retiring some equipment early cost it more, because writing an asset off sooner means a heavier charge in the year you do it. The decision reduced operating income by about USD 1.3 billion, operating income meaning the profit the business earns from actually running itself, before investment gains and tax. That is roughly 3% of the annual profit of AWS, Amazon’s cloud division, given up voluntarily by deciding its machines would wear out sooner.

Meta went the other way, extending certain servers to 5.5 years, which its earnings release says reduced 2025 depreciation by approximately USD 2.9 billion.

Follow that to the number shares are priced on: profit per share. Meta reported USD 23.49 for 2025, a shade below the year before. At the higher depreciation charge of the previous year, profit per share would have been about 79 cents lower. Whether the market noticed, nobody can say, because the same day’s results also carried revenue, growth and spending guidance. Anyone who hands you a share price movement attributable to the depreciation line is inventing it.

Both Meta and Amazon’s decisions were permitted and audited. Both sit in the notes to accounts, which is where this kind of judgement lives and where almost nobody looks.

Standards leave useful life to management deliberately, because management knows its own equipment. That is a sensible rule for a bakery. It becomes a very large lever when the assets run to hundreds of billions of dollars and the party making the estimate is the party whose profits it flatters.

A fair objection: depreciation is not cash. Whichever life you pick, the same money left the bank when you bought the oven. True, and it misses what the estimate touches. Useful life sets reported profit, which sets profit per share, which sets the price the shares trade at, which sets executive pay.

It also changes what the spending means:

• Buy a second oven while the first one still works, and you can bake twice as much bread. That is growth.

• Buy a second oven because the first one died, and you bake exactly as much bread as before. That is the cost of staying open.

If the chips really last 3 years rather than 6, a large part of the AI giants’ enormous spending is money they have to spend every year just to keep serving the customers they already have. Same bill, much smaller business.

Off-book deals:

One more bakery move, and this is the one that should make you sit up. The biggest ovens are not in the books at all.

Suppose the new oven is enormous and you would rather not show that debt. So you set it up differently. Your cousin takes out the loan and buys it. He owns 80% of the little company that holds it, you own 20%. You run the kitchen, you are the only person who ever uses that oven, and you pay him rent. The oven is not in your books. The loan is not in your books. Your accounts show a rent bill and nothing else.

This is the exact mechanism in play with Meta’s Hyperion data centre in Louisiana, at a scale of about USD 27 billion.

Meta and funds managed by Blue Owl Capital formed a Joint Centure, a separate company owned by both. Blue Owl’s funds own 80% and appoint its board. Meta owns 20%, provides construction and property management, and leases the finished campus back. Because Meta does not control the venture, it does not have to combine the venture’s accounts with its own, a process called consolidation. The debt and the data centre never touch Meta’s balance sheet, the list of what a company owns and owes. What would have been a capital cost becomes a rent bill.

Now on to the audit aspect of things. Every listed company’s auditor must publish what are called critical audit matters: the issues in the audit that were hardest and most subjective. It is one of very few places in a set of accounts where a professional is allowed to say that something troubled them, and the permitted vocabulary is narrow.

Ernst & Young, auditing Meta, wrote that assessing whether Meta has the power to direct the activities of the Blue Owl venture was challenging.

Ernst & Young signed the accounts. The treatment is permitted, the disclosure is proper, and they stood behind it. And then, in a public filing, in the one paragraph where the form allows it, they told every reader who knows how to listen that this call was close. Auditors do not spend that word casually.

One analyst, Ben Butler of Veritas Investment Research, has taken the other side outright, arguing Meta should be consolidating, because being sole tenant, equity investor and property manager together look a great deal like control.

The scale is why it matters. A study by Nikkei (Japan’s largest financial newspaper) published in July 2026 estimated the AI-related debt that Alphabet, Microsoft, Amazon, Meta and Oracle keep off their books at about USD 1.65 trillion. The debt those same 5 companies report on their own balance sheets comes to USD 1.35 trillion.

There is more borrowing outside the accounts than inside them.

There is a second route to the same destination, and OpenAI takes it. Where Meta part-owns the venture that holds the building, OpenAI does not own its data centres at all. It never bought the oven. It rents kitchen time from several different bakeries, each of which borrowed to build. Those bakeries are real companies with real debt.

• Oracle has issued USD 18 billion of bonds to finance the sites it is building for OpenAI.

• CoreWeave, an unrelated company that rents out AI computing power by the hour, has borrowed over USD 10 billion to build capacity for the same customer.

They compete with each other for OpenAI’s business rather than belonging to one group. Sherwood News put the total debt attached to OpenAI-linked projects near USD 100 billion. The buildings, the depreciation and the borrowing all sit on the suppliers’ books. That is how a company committing to more than a trillion dollars of infrastructure keeps a light balance sheet of its own. Nobody here is doing anything illegal. That is rather the point.

Circularity: the flour supplier lent you the money to buy his flour

The estimates flatter the cost side. A separate arrangement flatters the revenue side.

Your flour supplier lends you the money for the oven. You use it to buy his flour. His sales go up. On paper you are his customer, and in cash terms he is funding himself. Then he takes a share in your bakery, your bakery gets valued higher, and he books that rising value as profit, without selling the share. Both halves of that are happening.

Microsoft and OpenAI

• Microsoft has put USD 13 billion into OpenAI (the maker of ChatGPT) and holds a stake now worth about USD 135 billion.

• OpenAI has committed to buy USD 250 billion of cloud computing from Microsoft over roughly 6 years.

• In 2025, OpenAI paid Microsoft USD 17.2 billion, according to audited accounts obtained by the writer Ed Zitron and verified independently by the Financial Times.

• OpenAI’s revenue that year was USD 13.07 billion. It paid its largest shareholder more than it earned.

That is possible for 2 reasons, and the first is the more interesting.

1. Microsoft’s investment was never mostly cash. Semafor reported in 2023 that much of the USD 10 billion commitment was held as computing capacity (Azure) rather than money handed over. OpenAI draws that capacity down by running its models on Azure. Microsoft records the revenue, and little or no money actually moves. In the bakery, the flour supplier hands you a voucher instead of a loan, then counts the voucher as a sale when you redeem it.

2. What OpenAI does pay in cash comes from investors, not customers. A company can spend far more than it earns for as long as new money keeps arriving, which is what OpenAI’s successive funding rounds have done.

Amazon, Anthropic and OpenAI

• Amazon’s USD 8 billion in Anthropic (the creator of Claude) went in as convertible notes, which are loans that turn into shares later.

• Amazon’s USD 50 billion for OpenAI is reportedly structured partly as compute credits, meaning vouchers to buy Amazon’s own services, with USD 35 billion of it conditional on OpenAI listing publicly or building an AI that matches human ability across most tasks.

Nvidia and OpenAI

• Nvidia’s USD 100 billion buys shares with no voting rights, released in stages as OpenAI installs Nvidia chips.

• In the quarter ended 26 April 2026, Nvidia earned USD 53.5 billion from selling chips, which is its operating profit. Its profit before tax was USD 69.9 billion. This second figure is USD 16.4 billion larger, and almost all of that gap is a single line called other income, worth USD 15.9 billion, which Nvidia attributes mainly to its shareholdings rising in value. Nothing was sold and no money arrived. Nvidia’s own cash flow statement, which tracks money actually moving, strips the whole USD 15.9 billion back out for that reason. That other income line is 23% of the quarter’s pre-tax profit.

• In the same 3 months Nvidia put USD 18.6 billion into private companies and infrastructure funds, some of them AI developers who may go on to buy its chips.

• It also holds USD 27.4 billion of investments it is currently not allowed to sell.

Acadian, an American investment manager, makes the fairest defence available against the “AI Bubble” case. Classic bubbles do their damage by selling shares to outsiders at rising prices, so that when the story ends the loss lands on whoever bought last. This is not the same situation. The money here moves between the balance sheets of a handful of enormous companies funding one another. .

The difficulty is who owns the companies at the other end of that money. OpenAI and Anthropic are private. No ordinary investor can buy a share in either, which is what Acadian’s argument rests on. Microsoft, Amazon, Nvidia and Meta are not private. They are listed companies, and they sit in almost every index fund and pension fund in the world. So nobody needs to sell the public a single share in the risky part of this for the public to end up carrying the risk. The public already owns the companies writing the cheques. You cannot buy the bakery. You own a piece of the flour supplier.

The industry needs 30 times its current revenue

Anthropic has agreed to pay SpaceX (Elon Musk’s rocket company, which now also sells computing power) close to USD 45 billion over 3 years, disclosed in SpaceX’s own listing documents. Buried in the same disclosure: either side can end it with 90 days’ notice. A USD 45 billion commitment cancellable in 90 days is a 90-day commitment with a large number printed on it.

Meanwhile the spending climbs.

David Cahn of Sequoia Capital estimates AI infrastructure spending of about USD 1.5 trillion in 2026 and calculates it needs roughly USD 3 trillion of annual revenue to make sense. Bain, the consultancy, puts the requirement at USD 2 trillion a year by 2030.

The industry currently earns about USD 100 billion a year. Cahn’s arithmetic asks for 30 times that. Bain’s target means AI revenue doubling every year for 4 consecutive years, which no technology business has ever sustained at this scale. And that gap assumes the depreciation estimates hold and prices hold.

In the late 1990s, equipment makers such as Lucent and Nortel lent money to small telecom startups so they could buy their equipment. The sales were booked as revenue and looked real. When the startups failed, the loans went unpaid and the revenue turned out to be an illusion the lenders had financed themselves. J.P. Morgan Asset Management argues today’s build is different in 3 ways:

1. It runs mostly on cash these companies already generate rather than borrowing.

2. It buys real physical assets rather than promises.

3. It earns money as it goes.

The Nikkei figures complicate the first point, since USD 1.65 trillion of AI-related debt sits outside the accounts against USD 1.35 trillion reported inside them. And an asset being physical tells you it can be depreciated. It does not tell you it will earn back what it cost.

You may already own a piece of this cake:

Cisco made the routers and switches the internet ran on. In March 2000 it overtook Microsoft as the most valuable company in the world, on the view that the internet would need a great deal of its equipment. That view was correct. The internet did need it, and Cisco still sells it today. The company was right. The price was wrong.

Someone who bought at the March 2000 high and simply held did not see the price return to what they paid until December 2025. Almost 26 years. That is the share price alone. Cisco began paying a dividend in 2011, and an investor who reinvested those got back to even sooner. The point is that buying a genuinely dominant company at a genuinely wrong price can cost you.

For an Indian reader the mutual fund route is narrow. SEBI (the Securities and Exchange Board of India) caps the entire mutual fund industry at USD 7 billion invested overseas, and that limit is close to exhausted, which is why many international funds have stopped taking fresh money. That cap is not the whole picture.

Under the Liberalised Remittance Scheme, any resident individual may send up to USD 250,000 a year abroad and buy American shares directly through a broking platform, entirely outside the mutual fund limit. There are also international ETFs under a separate cap, and India-listed funds holding multinational subsidiaries. Indian household exposure is wider than the cap suggests, and I have no reliable figure for the total.

If you bought a Nasdaq fund for international diversification, you did not diversify. You bought a concentrated position in 7 companies. And if you hold those shares directly, this applies to you with more force, not less.

4 things to watch:

You do not need a view on whether AI is a bubble. You need to know which numbers would tell you.

1. Useful life. Disclosed in the notes to the accounts. Watch whether it is extended or shortened, and which way profits move.

2. Residual value. Not separately disclosed, which is the point. If a company starts talking up a healthy resale market for used computing power, ask what it now assumes its old chips are worth.

3. Depreciation against AI revenue. If depreciation grows faster than the revenue supporting it, last month’s milestone reverses.

4. Whether commitments can be cancelled. Termination clauses sit in the filings. A number you can walk away from in 90 days is not the number in the press release.

In summary:

Having read the filings, here is what the pattern looks like to me. The profits these companies report are flattered in 5 places, and every one is a judgement rather than a fact.

1. The cost of the machines is spread over a life they choose themselves, and the evidence says that life is too long.

2. Underneath that sits a second assumption about what the machines will fetch at the end, which moves the same number and which nobody discloses.

3. Part of the revenue comes from customers they funded.

4. Part of the profit is the rising value of stakes in those same customers, which is not money until it is sold.

5. Roughly USD 1.65 trillion of borrowing has been arranged so that it does not appear on their balance sheets at all.

All of it is disclosed. Every piece lives in the notes, the footnotes and the critical audit matters.

Put bluntly: the reported profits of the AI giants are softer than they look, the spending needed to justify them is around 30 times what the industry currently earns, and 7 of these companies are close to 40% of the Nasdaq 100.

There is a popular defence and it deserves an answer. It runs that even if AI software becomes cheap and interchangeable, the real value settles into the physical things underneath: the chips, the buildings, the electricity. That may well be right about where the value ends up. It does not help.

Those physical things are exactly what is being written off over 5 to 6 years, on estimates their owners make themselves. If the value moves to the machines, then how the machines are accounted for becomes the whole argument rather than a footnote to it.

So here is what I am actually telling you.

A reported profit is an assertion. It sits at the foot of a statement looking like a measured quantity, the way a bank balance does, and it is nothing of the sort. It is built on estimates a company made about its own machines. Every one of those estimates could reasonably have been made differently, and several of them were made differently by the company’s closest competitor in the very same month. Nowhere is that gap wider today than in artificial intelligence, where a single view about how long a chip lasts moves billions of dollars and appears nowhere in the headline.

The more useful thing from all this is the habit. Every set of accounts you will ever open, for any company in any industry, contains 2 kinds of number.

1. Some were counted. Cash in the bank, units sold, wages paid.

2. Others were estimated: how long the machinery lasts, what it will fetch at the end, which debts will go bad, what a stake in another company is worth today.

Reading the published accounts will not tell you which is which. They print in the same font and add into the same total. You have to ask. Which of these figures is an estimate, and who made it. You do not need to be an accountant to ask either one.

For the record: I am not calling a bubble. I do not know whether these estimates prove wrong. If they do, the correction arrives as a write-down, the moment a company admits its assets are worth less than its books have been claiming. The bakery owner accepting the oven is finished, at the scale of a trillion dollars.

What would change my mind:

1. A sustained period in which AI revenue growth outpaces depreciation growth on shortened useful lives.

2. Disclosure separating revenue earned from investors and their affiliates from revenue earned from everyone else.

Neither exists yet. Both are within these companies’ power to provide.

One more assumption holds all of this up, and nobody states it out loud: that what these machines produce will hold its price for 6 years. It has not held its price for 6 months. That is Part 2.

The headlines tell you who is building the future. The accounts tell you who is paying for it, and on what assumption. That assumption is the omission.

Subscribe now

Share

Leave a comment

Sources:

Primary filings and company disclosures: Amazon Q4 2024 earnings call (6 February 2025) and Form 10-K, for the useful-life change from 6 years to 5, the stated reason, and the USD 700 million, USD 600 million and USD 920 million effects; Amazon Q4 2025 earnings release for AWS operating income of USD 45.6 billion; Amazon Form 10-Q disclosures on the Anthropic convertible notes; Meta Q4 2024 earnings release (29 January 2025) and 2025 Form 10-K for the extension to 5.5 years and the approximately USD 2.9 billion effect; Meta Q4 2025 earnings release for diluted EPS of USD 23.49 against USD 23.86 in 2024; Meta press release of 21 October 2025 on the Hyperion joint venture, the 80/20 split and the approximately USD 27 billion development cost; Ernst & Young’s critical audit matter on control of that venture; NVIDIA Form 10-Q for the quarter ended 26 April 2026; SpaceX Form S-1 (May 2026).

Reporting and analysis: Bloomberg on AI revenue exceeding depreciation, drawing on Exponential View (June 2026); Bloomberg Law on the accounting treatment of AI data centre joint ventures, including Veritas Investment Research’s dissent; Ed Zitron on OpenAI’s audited 2025 accounts, verified by the Financial Times, for the USD 17.2 billion paid to Microsoft, revenue of USD 13.07 billion and total costs of USD 34 billion; Semafor (2023) on the composition of Microsoft’s USD 10 billion commitment as largely Azure capacity rather than cash; CNBC on Michael Burry’s depreciation analysis (11 November 2025); Nikkei Asia on off-balance-sheet AI debt (July 2026); Sherwood News on OpenAI-linked project debt; reporting on the Amazon and OpenAI investment structure, including the compute-credit component; Tomasz Tunguz’s compilation of OpenAI’s infrastructure commitments.

Counter-arguments: J.P. Morgan Asset Management, “Does circularity in AI deals warn of a bubble?” (October 2025); Acadian Asset Management, “Straight Talk About Circular Deals in AI” (April 2026).

Estimates and own arithmetic: David Cahn, Sequoia Capital, on required AI revenue; Bain & Company on USD 2 trillion by 2030. The 30-fold gap and the doubling calculation are mine, against those published estimates. The residual value argument is mine; I make no claim that any company has revised that assumption, only that the lever exists and is not separately disclosed.

Figures below are for Tax Year 2026-27 and assume the new tax regime, which is the default. Assumptions are stated wherever a number appears.

They told you the number. You did the arithmetic in your head, divided by 12, and felt a certain satisfaction. Then the payslip arrived and something felt different..

CTC (cost to company, the total annual amount your employer spends on employing you) includes several things that never touch your bank account:

• The Employer’s Provident Fund Contribution

• Gratuity provision

• Medical Insurance premiums

• Any variable pay you have not yet earned

What lands in your account is what remains after all of that, net of the following additional items:

• Your own PF deduction

• Professional Tax (INR 2500 / year)

• TDS (tax deducted at source, the income tax your employer withholds each month before paying you)

Here is the current arithmetic. Assume basic salary at 50% of CTC, employer PF at 12% of basic, gratuity provision at 4.81% of basic, employee PF at 12% of basic, professional tax of 2,500 a year, the standard deduction of 75,000, and the new regime slabs for Tax Year 2026-27.

At a CTC of INR 12 lakh, the gap runs to about INR 1,75,000 a year, and 99% of it is Provident Fund across both contributions plus the gratuity provision. Income tax takes away nothing at all, because the rebate under the new regime cancels your liability entirely up to INR 12 lakh of taxable income.

At INR 24 lakh, the gap is about 5,88,000. PF and gratuity account for 59% of that. Income tax takes the other 41%. You will see this again when you file your tax returns.

So the gap does not simply get bigger. It changes composition. Lower down it is deferred pay and you get it back eventually. Higher up, a growing share is tax, and that share does not come back.

Run your own version of this table with your actual structure. Then budget on the in-hand figure. CTC is a number for HR conversations.

The variable component

Most promoted packages carry a variable slice. It is real but it is not guaranteed, and those are different planning problems. If 20% of your CTC is variable, build your baseline on the 80% you are certain to receive. When the variable arrives, route it to investments or a specific goal before it has time to become a lifestyle. Treat it as a windfall even though you earned it.

Why your basic salary suddenly matters more

Basic salary drives the following:

• Your PF contribution

• Your gratuity calculation

• Your HRA exemption ceiling

• Your leave encashment

For years, most Indian companies kept basic at 30% to 40% of CTC and loaded the rest into allowances, which minimised statutory contributions and maximised take-home. The four labour codes came into force on 21 November 2025 and changed the mechanics.

Under the Code on Wages, 2019, “wages” means basic pay plus dearness allowance plus retaining allowance, and the allowances excluded from that definition cannot exceed 50% of total remuneration. Any excess gets added back into wages for statutory purposes. In practice this pushes most salary structures up to a 50% basic floor.

Two pieces of precision matter here, and most explainers skip both. The rule bites on wages as a share of total remuneration, not on basic as a share of CTC, and those are not identical, because CTC includes employer contributions that sit outside remuneration. And enforcement is uneven. Labour sits on the Concurrent List, central rules were still being finalised through early 2026, and states have notified their own rules at different speeds. Maharashtra, Karnataka, Gujarat, Haryana and Madhya Pradesh moved early. Others have not. Your employer may have restructured, or may be waiting.

If your CTC has restructured, expect your monthly take-home to fall by roughly INR 2,000 to INR 4,000 at a CTC between INR 12 to 18 lakh. On a INR 12 lakh package moving from 40% basic to 50%, the drop works out to about INR 2,880 a month, and about INR 28,800 a year extra flows into your PF account across both contributions. That is money moving from your wallet into your retirement corpus. It is still yours. You will simply see it later with interest.

The gratuity rule almost everyone gets wrong

You will read, in a great many places, that gratuity vests at exactly 5 years and that leaving at 4 years, 11 months and 29 days forfeits the lot. The statute is more generous than the folklore.

Section 4(1) of the Payment of Gratuity Act, 1972 requires 5 years of continuous service. Section 2A then defines continuous service, and deems an employee who has worked 240 days in a 12-month period to have completed a year. The Madras High Court applied exactly this reasoning in Mettur Beardsell Ltd. v. Regional Labour Commissioner (1998 LLR 1072), holding that 240 days worked in the fifth year completes the 5th year for gratuity purposes. Delhi, Kerala and Punjab & Haryana High Courts have taken similar views.

So, the working rule is 4 years plus 240 days in the 5th year, not a full 60 months.

The caveat: there is no Supreme Court ruling squarely on this point for gratuity, so these judgments are persuasive rather than universally binding. Some employers pay without argument. Others require you to file before the Controlling Authority. Know which kind you work for before you plan your exit date.

The amount, when it vests, is 15/26 of your last drawn basic for every completed year. On a basic of INR 50,000 a month with 5 completed years, that is 1,44,231. The statutory ceiling is INR 20 lakh on which you do not pay any income taxes in the course of your professional lifetime.

One more thing the codes changed: fixed-term contract employees now qualify for gratuity after 1 year of service, rather than 5. If you are on a fixed-term contract, check this. It is new, and payroll teams are still catching up.

If your promotion came with equity

ESOPs (employee stock option plans, the right to buy company shares at a fixed price later) are taxed at two separate moments, and people are routinely surprised by the first one.

At exercise, when you actually buy the shares, the difference between the fair market value on that date and your exercise price is treated as a perquisite and taxed as salary income at your slab rate. You pay tax on a gain you have not converted to cash.

At sale, the difference between your sale price and that same fair market value is taxed again, as capital gains.

If your employer is an eligible start-up under Section 80-IAC, the TDS on the perquisite can be deferred by up to 48 months. Most employers are not eligible. Ask rather than assume. RSUs (restricted stock units, shares granted outright rather than optioned) work similarly, taxed as salary on the vesting date at market value. Do not count unvested equity as current wealth. It belongs to a version of you who has not turned up yet.

What to do right now:

• Get the full breakup from HR in writing: fixed salary, variable percentage and how it is measured, every allowance itemised, and the employer PF and gratuity provisions.

• Check whether your structure has been realigned to the 50% wage floor, and whether your state has notified its rules. If your basic is still at 35%, you now know a change is coming.

• Build your own CTC-to-in-hand table using your actual figures. Budget on the in-hand line.

• Confirm your completed years of service and your day count in the current year. If you are close to the 240-day mark in your fifth year, that is a date worth knowing before you take any call about leaving.

• If your package includes ESOPs or RSUs, get the vesting schedule, exercise price, exercise window after exit, and the tax treatment in writing.

Part of the First Promotion Playbook.

Nobody upgrades their life in one dramatic decision. There is no morning where you wake up and choose to spend your entire raise.

Subscribe now

Instead, there are twelve small decisions, each one perfectly reasonable.

A slightly better flat, because you can afford it now. You stop checking auto prices on the app, because comparing fares feels beneath your new salary. Weekend brunches that were occasional treats become weekly defaults. A gym upgrade, because the new one is closer to the office. Two subscriptions you will forget to cancel by month three.

None of these feel like financial mistakes when you make them. That is the design. They arrive one at a time, each dressed as something you have earned, and together they consume a raise before it has a single chance to compound.

Run the numbers on how fast it happens. You receive a INR 10,000 monthly raise after tax. Rent goes up INR 3,000. Food spending climbs INR 2,000. The gym costs INR 1,500 more. Weekends rise INR 2,000. Cabs replace autos for another INR 1,500.

Total lifestyle increase: INR 10,000. Wealth created: zero.

You were never careless. No single decision was wrong. The raise simply arrived without a plan, and money without a plan follows the path of least resistance, which always leads through your spending.

Now run the numbers the other way. That same INR 10,000 per month, invested at 12% for 20 years, becomes INR 99 lakh. A 2BHK down payment in most Indian cities. That is what the twelve small decisions actually cost. A market crash could not have taken that much from you. Ubers, brunches, and a gym you visited twice a week did.

Lifestyle inflation does not feel like a INR 99 lakh mistake. That is exactly why it is one.

The reason it catches nearly everyone is that it works on three levels at once. Socially: when your income grows, the people around you recalibrate what you can afford, and you start benchmarking against a richer peer group without noticing you switched. Identity: a new title feels like it should come with visible markers. Relief: you worked hard for years, and spending finally feels like acknowledgement.

All three are deeply human. None of them build wealth.

And beneath all three sits the mechanism most financial advice never mentions: the ratchet effect. Lifestyle upgrades are easy to add and psychologically brutal to reverse. Once you have the bigger flat, moving back to the smaller one feels like failure, even though you lived there happily three months ago. Nothing about the smaller flat changed. Your reference point did. Upgrades move in one direction, and every one you add becomes the new floor you cannot imagine dropping below.

That asymmetry is the real trap. You are never deciding “should I spend more this month.” You are deciding “what will I be unable to walk back from next year.”

There is a rule that beats all of this, and it is simple enough to carry in one line: the First Claim Rule. Your future gets the first claim on every increment. Not the last.

Before your first promoted salary hits your account, decide what percentage of the increase goes to savings and investments. What goes first, before any spending begins. Even a 50-50 split, half to your future, half to your present, changes the trajectory of a decade. The timing is the entire trick: the decision must happen before the money arrives, because after it arrives, it has already been mentally spent.

Watch the financially calm people you know and you will see the rule running underneath their lives. They are not depriving themselves. They are sequencing. Increment arrives. Savings rate rises first. SIP increase is already set. Emergency fund top-up is scheduled. Then, and only then, one deliberate lifestyle upgrade with what remains. The upgrade still happens. It just never happens at the expense of the future.

Your salary is the input. Your wealth is the output. What separates them is what you do in the week each raise arrives.

Twelve small decisions will be made either way. The only question is whether you make them or they make themselves.

What to do right now:

1. Calculate your exact increment after tax. The actual additional rupees hitting your bank account each month. The gross number HR announced is for congratulations, and budgets built on it fail by exactly the tax difference.

2. Apply the First Claim Rule before the first promoted salary arrives: decide what percentage of the increment goes to savings and investments. Even 50% changes the decade.

3. Increase your SIP before the first promoted month ends. Log in. Change the number. It takes four minutes, and the SIP that is not increased in month one rarely gets increased at all.

4. Choose one deliberate lifestyle upgrade with what remains. One that you pick consciously, instead of twelve that pick themselves.

5. Set a 90-day review reminder. Check whether the split you decided is holding, or whether the ratchet has already added a new floor you did not agree to.

Subscribe now

Share

Leave a comment