AI compute could become more expensive even as GPUs improve

Supported by Bright Data

Turn the Web Into A Data Pipeline. Generate Your API, in Minutes.

If your team is still patching selectors and rotating proxies, you’re maintaining infrastructure your competitors stopped owning.

Scraper Studio - launching today from Bright Data turns any public website into a hosted data API from a single prompt:

• Self-healing scrapers that detect site changes and patch themselves

• Proxies, unblocking, and parsing handled - not your problem

• Build from Claude Code, Cursor, or Codex via CLI - or visually in the control panel

• Scale production data pipelines without owning the stack

• 5,000 free credits every month

No sales call. No demo. Just spin up a scraper and watch it run.

Start free →

For most of computing history, better technology meant lower costs.

Processors become faster, storage becomes cheaper, networks carry more data and software extracts more useful work from each machine. The thing is, even when the latest hardware is expensive, the cost of completing standard (standard — if you're not familiar) computing tasks generally decreases.

But, quantum leap developments in the field of artificial intelligence may produce more complex results.

The costs of creating tokens, running models, or completing individual AI tasks may continue to decline while the market price of strategically important computing increases. Plain and simple (at least in most cases), this could happen if AI systems become economically productive faster than chip factories, memory suppliers, power grids, and data centers expand.

A company that right now spends $200,000 or more annually on engineers might rationally spend a large portion of that amount on AI systems.

In such a scenario, computing is no longer valued simply as rental hardware. Now here's the interesting part: this'll be a rare input capable of producing valuable intellectual work (and this is key).

This could support prices several times higher than current GPU (Graphics Processing Unit) rental rates. But achieving tenfold improvement on a long-lasting basis requires more than AI capabilities. In practice, this requires supply constraints, limited competition, and workloads that are valuable enough to absorb the higher costs. Which, when you think about GPU, makes perfect sense.

Testing the $250,000 per year calculation

A computing resource generating annual rental income of $250,000 should be priced approximately:

$250,000 \div 8,760 = $28.54 \text{ per hour}

Assuming the hardware is rented hourly this year.

Real-world utilization is actually lower. From what we can tell, at 70% billable utilization, (for the most part) the price increases to about $40.77 per hour. Real talk, maintenance, failures, spare capacity, customer turnover — and periods of weak demand all reduce the number of hours an infrastructure or what's known as an infrastructure provider can make money from.

As of August 2026, Lambda lists committed H100 cluster capacity at approximately $5.54 to $6.16 per GPU (Graphics Processing Unit) hour, depending on cluster size. At the risk of stating the obvious, that's about $48, (which is pretty standard)500 to $54,000 per year if used continuously. Some interruptible or marketable H100 offerings are much cheaper, although they don't provide the availability, network — or assurance required for mission-critical autonomous agents.

At an enterprise-level committed capacity comparison, an increase from about $50,000 to $250,000 per GPU-year would mean an increase of about fivefold, not fifteenfold.

Ten- or fifteen-fold comparisons become possible only if the starting point is deeply discounted, interruptible, or wholesale GPU prices. As it turns out (which makes a lot of sense when you think about it), here's the thing, this isn't necessarily an appropriate basis for AI workers who are basically expected to continue operating within production systems.

Workforce comparisons also need qualification. Essentially (for what it's worth), the average salary of a software developer in the US in May 2025 was approximately $135,980, while the highest paid 10% earned more than $214,670. Compensation at leading tech companies can be higher when bonuses and equity are basically included, but $250,000 represents experienced or highly paid engineers, not the entire software development market.

The broader claim remains economically important: AI systems capable of replacing high-value engineering jobs could create far greater infrastructure which is actually essentially an infrastructure spending than typical AI (Artificial Intelligence) applications today.

But the engineer's value doesn't automatically transfer to the GPU provider.

GPUs are incapable of capturing the full value of an AI worker

Suppose a company spends $250,000 per year on an engineer and gets access to an AI agent with similar capabilities. Companies can't necessarily afford to spend the entire $250,000 on raw computing.

You may still need to pay for:

• Base model or license model

• Storage, networking, and databases

• Sandbox development environment

• Security and identity controls

• Human review and escalation

• Observation and evaluation systems

• Failed attempts and repeated executions

• Integration with internal tools

• Responsibility for incorrect or unsafe output

The system can also automate only part of the engineer's work. Writing code is one component of software engineering. Engineers also clarify requirements, (usually) negotiate tradeoffs, coordinate with other teams, investigate production incidents. At the risk of stating the obvious, which, when you think about it, makes perfect sense. Understand organizational constraints and Accept responsibility for decisions.

Theoretical value can still quickly shrink.

Assume the agency targets jobs related to employees worth $250,000. Now here's the interesting part: it covers 80% of the person's tasks and performs those tasks reliably 85% of the time (if you think about it).

The gross labor value is: