Who Profits When AI "Saves" Developer Costs

The pitch counts the engineers you fire. It forgets the GPUs you buy.

The pitch

You've heard it. AI writes the code, so you need fewer developers, so you save money. Developers get raises and quit. Silicon is a clean line item.

Count only the salaries and it sounds airtight. But a balance sheet has two sides, and the pitch reads only one. It's a cost shell game.

Why the math feels right (and isn't)

Start small. If you run a small business, software feels like a one-time purchase. You pay once to build the app. Then the recurring bill arrives.

Hosting alone is cheap in India: roughly ₹2,000–₹5,000 a year for shared hosting, ₹6,000–₹20,000 for cloud or VPS, and ₹80,000+ for dedicated servers. The real cost is upkeep. For a custom web app built for ₹5–30 lakh, hosting runs ₹2,000–₹25,000 a month and maintenance runs 15–20% of the build cost, every year. That's roughly ₹1–9 lakh a year, forever.

Big Tech scaled that logic to the enterprise, and it breaks there. Enterprise software is never finished. New regulations, integrations, attack surfaces and customer demands keep arriving. You don't buy it. You feed it.

The side they skip: Microsoft's hardware bill

Look at the company that told the world AI would transform software development.

• For fiscal 2026, Microsoft's capital expenditures and finance leases reached about $175 billion, with roughly the same guided for fiscal 2027. About two-thirds went to short-lived assets like CPUs and GPUs.
• Cash property and equipment spending rose nearly 80% to about $116 billion. About $25 billion of the increase came from higher memory and component prices, not extra capacity. Microsoft paid more for the same boxes.
• Amazon, Microsoft, Alphabet and Meta plan to spend up to $725 billion combined on capital projects in 2026, a 77% jump.
• Analysts note Microsoft has gone from capital-light and high-margin to capital-intensive in a few years, and gross margin has compressed for three straight quarters.

Who actually supplies the hardware

The money doesn't vanish. It lands with a short list of suppliers:

• Nvidia: IDC puts it at roughly 81% of AI data center chip revenue in 2026, with $75.2 billion of data center revenue in one quarter. Hyperscalers like Microsoft account for about half. Microsoft deployed the first large-scale cluster of Nvidia GB300 GPUs.
• AMD: a meaningful second source inside Microsoft and Meta.
• Broadcom: networking and custom chip solutions.
• Memory makers (SK hynix, Samsung, Micron): SK hynix alone holds more than half of the high-bandwidth memory market. This is the part whose price spiked.
• GPU "neoclouds" and Microsoft's own Maia chips fill the rest.

Every dollar of "AI savings" that doesn't reach shareholders reaches these companies.

Who sold the idea

No single person invented it, but look at who said it loudest:

• Anthropic's Dario Amodei said AI would write 90% of code within three to six months.
• Mark Zuckerberg said Meta was building AI that codes at a mid-level engineer's level.
• Satya Nadella said 20–30% of Microsoft's code was written by software.
• Nvidia's Jensen Huang advised kids not to learn to code.
• Salesforce's Marc Benioff suggested the company might hire no more software engineers.

Notice the pattern. These are AI labs, chip makers and cloud sellers: the people who profit when you buy compute. That is a conflict of interest, not proof of bad faith. And the big prediction hasn't landed: industry-wide transformation hasn't materialized the way Amodei forecast. (Disclosure: Anthropic, which makes this assistant, is one of the companies named.)

What developers actually cost

Proponents claim an avg 100 engineering team burns ₹20–50 crores a year for enterprise. Check it against the market. Current India data puts entry-level engineers at about ₹4–10 lakh, mid-level at ₹12–22 lakh, and seniors at ₹25–40 lakh a year. IT services tech leads and architects with 8–12 years mostly land at ₹18–32 lakh.

Build a realistic team of 100:

• 40 juniors × ~₹6L = ₹2.4 Cr
• 40 mid-level × ~₹15L = ₹6 Cr
• 20 seniors × ~₹30L = ₹6 Cr

Total: about ₹12–15 crores a year. Now "transform with AI." Keep 10 elite architects at ₹30 lakh each. Payroll drops to ₹3 crores. The board applauds.

Then someone has to buy and run the hardware.

Do the math they didn't

Microsoft's product R&D headcount fell by 3,000 to 77,000, and total headcount fell by 5,000 to 223,000. The CFO said it would fall again in fiscal 2027. Now price it. A Microsoft software engineer posting lists base pay of roughly $102,000–$235,000. Assume a generous $200,000 fully loaded (my estimate).

• 3,000 fewer R&D engineers ≈ $0.6 billion a year.
• 5,000 fewer employees overall ≈ $1 billion a year.
• Meanwhile, cash property and equipment spending rose by roughly $50 billion.

The headcount savings are a rounding error next to the hardware increase. At hyperscale, compute now outweighs developer pay. Lower developer cost isn't funding the AI buildout. It's a drop in the bucket.

The bill nobody shows you: inference

Every user interaction, code generation and agent loop costs money. Forever. And here's the paradox: token prices keep falling, yet bills keep rising. Cost per token at a fixed capability level drops roughly 5–10× a year, but enterprise AI compute spend still rose 320% over the same period. Cheaper unit, bigger invoice.

• A McKinsey FinOps survey (May 2026) found 93% of enterprises are exceeding their AI budgets.
• About 60% of agentic AI spend goes to refining responses, not the first call.
• The FinOps Foundation places 80–90% of AI spending in inference.
• Harness estimates 26% of enterprise AI spend is wasted.
• Uber's 5,000-strong engineering team reportedly burned its annual AI budget in four months.

So what do 10 million interactions a month really cost?

I rebuilt it from published API prices. One interaction = one call of ~2,000 input and 500 output tokens. Exchange rate assumed at ~₹88/$.

A mid-tier model with some caching or routing lands near ₹9 crores a year (about ₹75 lakh a month). That is the number the pitch quietly leaves out. Agentic workflows make it worse, because they hit a model 10–20 times per task.

The balance sheet

Traditional: ₹13 Cr. AI, as sold: ₹13 Cr. Net savings: zero. You didn't eliminate the cost. You moved it from engineers to cloud and silicon providers.

Pick a premium model and let agents loop, and you're at ₹18 Cr+, paying more than the team you let go. Cheap open models can cut the bill to about ₹5.4 Cr, but that saving doesn't come from firing people. It comes from measuring spend, routing work to cheaper models, capping agent loops and reviewing output. All of that takes skilled humans.

*Enterprise hosting of ₹1 Cr is an illustrative assumption, not a benchmark.

So why does nobody talk about it?

• Salaries are visible. Depreciation is quiet. A layoff makes a headline. GPUs go on the balance sheet and get written off over years. (My analysis.)
• Hosting never gets a champion. Nobody is promoted for noticing the electricity bill, and the cost sits in a different department from the savings claim. (My analysis.)
• Layoffs sound like efficiency. One commentator argues layoffs are really the financing for capex. Treat that with care: Deutsche Bank analysts flag "AI redundancy washing," where firms blame AI for cuts they'd have made anyway, and Bloomberg data suggests about half of AI-attributed layoffs end with the roles rehired offshore or at lower pay.

Investors do notice. Microsoft expects to stay capacity-constrained through the rest of 2026 even after all that spending. What's missing is that the developer-replacement pitch itself rarely includes the hardware.

The collateral damage

You burn institutional knowledge

The developer who knows why that odd billing rule exists keeps it in her head. Fire her and it walks out the door. No model can retrieve what was never written down.

You buy technical debt at scale

Veracode found 45% of AI-generated code samples introduce OWASP Top 10 vulnerabilities, with cross-site scripting in 86% of samples tested. CodeRabbit found AI-generated pull requests average 10.8 issues versus 6.4 for human code. GitClear saw code duplication spike 4×. Ten architects can't review that flood.

You kill the junior pipeline

Seniors are promoted juniors. Stanford's Digital Economy Lab reports employment for 22–25 year-olds in AI-exposed jobs about 19% below trend, with no comparable drop for experienced workers. Indeed's Hiring Lab shows senior postings up 13.5% since January 2025 while entry-level postings fell 6.3%. If nobody hires juniors today, who are your senior architects in 2033?

And Gartner projects over 40% of agentic AI projects will be cancelled by the end of 2027, citing cost, unclear value and weak risk controls.

What to do instead

• Count both sides. Every AI business case must include compute, hosting, power and review cost next to the salaries removed.
• Ask who's selling. Treat "AI replaces developers" claims from compute sellers as marketing until you see the full ledger.
• Augment, don't replace. Give every engineer an AI copilot. Measure output per engineer, not headcount cut.
• Use cheaper models where you can. Route simple work to smaller or open-weight models like Llama. One enterprise AI company reportedly cut inference costs sixfold after leaving closed per-token APIs.
• Build a cost-to-value veto. Every AI feature needs a unit cost per useful task before launch. Cap agent loop depth. Track spend per workflow.
• Reinvest in human QA and security, and keep hiring juniors to review AI output.

The takeaway

AI is a powerful tool. It is not a free lunch. The money doesn't disappear when you fire engineers. It moves to Nvidia, the memory makers and the cloud providers. The honest pitch is "make 100 people far more productive, and spend compute where it earns its keep."

Anyone promising a saving with no hardware line on the sheet is sending you the bill in a different envelope. Read both sides before you sign.