In several of our recent interviews with senior engineers and engineering managers, I heard the same argument:
“Tokens are heavily subsidized. LLM usage will eventually collapse, and software engineering jobs will come back.”
After hearing this in three separate interviews on Friday, here is how we currently see the market.
1. Estimates of token subsidization vary significantly, ranging from 10x to as much as 25x in more extreme analyses. The most common forecast we see is around 12x. That is the situation today.
2. With the $100/mo Max license on Claude Code or Codex, the average usage can easily represent more than $1,200, and sometimes more than $2,000, in underlying compute costs.
It may be difficult to calculate precisely on a daily or weekly basis, but there are certainly days when a few hours of work generate solutions that would previously have cost five figures. And this can happen repeatedly.
3. Top performers no longer need to delegate every moderately complex task to teams of 5, 10, or 20 people.
Some of these tasks can now be completed individually. Others require only two or three additional people.
Beyond the communication overhead, meetings, briefs, handovers, and reviews, traditional team structures also introduce the human factors of vacations, turnover, recruitment, onboarding, and performance variability.
All of these carry mental, operational, financial, and time-related costs, as well as additional risk.
4. There will always be software engineers.
But companies will no longer hire almost anyone who has completed an eight-week bootcamp or online course.
The nature of the work no longer creates the same pressure to fill seats when individual productivity is increasing so dramatically.
5. The work itself is also changing.
A significant part of a developer’s job is becoming the management of agents or small automated teams. It increasingly includes product management, architecture, horizontal provisioning, systems access management, QA process ownership, risk management, and failover planning.
Capable developers are developing senior-level skills faster because they can iterate more frequently and gain broader hands-on experience.
6. Working with AI is fundamentally a trial-and-error process.
It requires daily experience with different models, tools, and systems built on top of language models.
Some models are strong in certain areas and weak in others. As new models, harnesses, and development environments are released, these capabilities also change.
A meaningful part of a developer’s experience will come from investing several thousand hours in working with the leading tools and environments.
They need to understand when a problem can be solved in two minutes with the right prompt, prepared tools, and relevant skills, and when the same approach is likely to create unreliable output or cause serious damage.
7. A lack of sufficient experience in point six is already becoming a decisive hiring factor.
I have been discussing this since last year.
Companies will find it difficult to justify spending six months teaching someone which tools work in which situations, how to use them effectively, and which skills or configurations are appropriate for Cursor, Claude Code, or similar environments.
This is becoming one of the core capabilities required to balance speed, stability, reliability, and security throughout the development process.
8. Tokens may be subsidized by 12x today, but the cost of energy, chips, and hardware will continue to decline.
History repeatedly shows that broader distribution reduces costs. Competition creates alternative solutions. A meaningful share of today’s pricing is driven by scarcity, particularly the limited availability of chips, data center capacity, and specialized hardware.
That capacity is being built rapidly.
We saw similar patterns with graphics cards during the crypto boom, fuel prices during wars, and the early cost of solar panels and electric vehicles. New infrastructure is expensive at first, but costs tend to normalize as production and competition increase.
9. When OpenAI and Anthropic eventually complete successful IPOs, millions of investors will own a piece of the market.
That will create even stronger incentives to integrate these systems across industries and make the economic model sustainable.
The services cannot remain prohibitively expensive for customers, but the companies providing them also cannot operate indefinitely at unsustainable losses.
10. We should not overlook Chinese models or self-hosted solutions running on DGX systems and other hardware.
Today, a machine that costs approximately $6,000 as a one-time investment and around €20 per month in electricity can already produce remarkable results.
Two or three such machines can support an entire team and provide substantially more capacity than the monthly token allocations available to many organizations.
The conclusion is not that software engineering disappears.
The conclusion is that the economics, team structures, required skills, and definition of a strong engineer are all changing at the same time.
