Alternative computing: the real cost of switching

Optical AI research reveals a practical startup question: how much flexibility can a customer trade for greater efficiency?

Alternative computing has to survive a very ordinary moment: a customer asking what happens when their needs change. A processor can perform an impressive calculation, yet leave its buyer struggling to update a model or integrate the result. For founders exploring alternatives to GPUs, that gap deserves as much attention as the breakthrough.

The strongest starting point is a task repeated so often that improving it changes what a customer can afford to offer. Optical image generation suggests how such an opportunity might emerge. It also exposes a question every ambitious hardware pitch should answer: what flexibility does the customer give up to get the promised efficiency?

Efficiency comes with conditions

In his optical diffusion presentation at YC Paper Club, starting at 22:23, Ilker Oguz describes using light to help generate images. Diffusion models progressively turn noise into an image. His team’s approach uses the propagation of light through optical layers to predict noise that can then be removed.

The accompanying research by Oguz and his coauthors adapts the generation process to reuse optical settings across repeated steps. Its energy comparison assumes passive optical layers. The experimental prototype uses a programmable device, so that assumption matters when interpreting the efficiency claim.

There is another boundary. In his explanation of adoption barriers from 19:27, Oguz identifies costs associated with converting information between electronic and optical forms, alongside programming and calibration. Light performing a calculation efficiently does not establish the cost of delivering a finished result.

My reading is that these constraints help define the first customer. Look for a workload whose repetition, stability and input requirements suit the machine. That gives the startup a concrete proposition to test, with a buyer who can explain why the improvement matters.

Put a price on changing your mind

Imagine a startup offering image generation for a retailer. Before considering specialized hardware, the team needs to understand how often customers require different output styles, higher resolution or a new model. A system optimized around a stable task could fit one service beautifully and create awkward restrictions for another.

Ask the supplier to demonstrate an update. Does changing the model require software, recalibration or replacement components? How much engineering work does it create, and can the service keep running during the change? These questions turn an abstract promise of flexibility into something a buyer can assess.

The same discipline applies to suppliers’ roadmaps. If a commercially essential feature is still experimental, include its absence in the current purchasing decision. A useful prototype can justify a research partnership while a production commitment needs evidence about support, delivery and repeatability.

For a small team, engineering time belongs in the calculation. Moving people away from customer requests to maintain a specialized system has a cost, even when the electricity bill improves. The saving has to survive the work around the calculation.

Measure the outcome the buyer values

Start a pilot with the customer’s acceptance criteria. For the hypothetical retailer, that could mean the cost of producing an image approved for use, including rejected outputs, retries and review. Counting every generated image equally would hide the work needed to turn a sample into something commercially useful.

Choose the comparison carefully, too. Test against the best practical version of the existing service, using equivalent quality requirements and realistic demand. Record quiet periods, busy periods and failures. A machine that looks attractive under continuous use may produce a different business case when orders arrive unevenly.

This is familiar territory for teams dealing with voice agent failures after the demo. A fast response is only one part of completing a customer’s request. Hardware evaluation deserves the same attention to the entire interaction, including recovery when something goes wrong.

Different products also need different measures. For a team developing AI hearing aids, a hardware trial might prioritize battery life, response time and performance in everyday environments. Those criteria would need their own evidence; an image-generation experiment cannot establish them.

Make switching a manageable experiment

Founders building applications can preserve room to experiment by keeping hardware-specific code separate from customer-facing features. Choose a bounded task for the first trial, retain a working fallback and define what would justify expanding it. That makes disappointing results useful without putting the whole service at risk.

Before changing processors, check whether computation actually causes the problem. A coding assistant producing outdated answers may need access to current information through search. Faster execution would leave the information gap intact. Diagnosing the constraint protects both the experiment’s budget and its credibility.

Hardware startups can help by selling a pilot with explicit responsibilities: who integrates the system, who measures quality and who handles updates. Agree on the conditions for a purchase before the experiment begins. A successful demonstration becomes more valuable when both sides know what commercial decision it supports.

The founder’s takeaway

Alternative computing deserves curiosity and a demanding first customer. The opportunity is to match an unusual technical strength with a task whose economics improve under real operating conditions. Repetition can make specialization valuable; frequent changes can make flexibility worth paying for.

Before committing, establish the workload, measure the complete result and test an update. The right first customer is the one whose constraints make your unusual hardware useful. Start there, and a research achievement has a clearer path to becoming a business.

Video: Y Combinator, What If We Stopped Using GPUs? | YC Paper Club, October 2, 2026.

The saving has to survive the work around the calculation.

The right first customer is the one whose constraints make your unusual hardware useful.

▶

Watch the original episode on YouTube: Y Combinator.

Sources

• Y Combinator — What If We Stopped Using GPUs? | YC Paper Club — Published October 2, 2026. Relevant chapters: 19:27 on optical computing adoption barriers; 22:23 on optical diffusion; 30:03 for the optical computing discussion.
• Optical Diffusion Models for Image Generation — Primary research by Ilker Oguz and coauthors. The methods and appendices distinguish passive-layer energy assumptions from the programmable experimental apparatus.
• Public transcript — What If We Stopped Using GPUs? | YC Paper Club — Transcript consulted for Oguz's explanations of conversion costs, fixed optical settings and experimental limitations. Technical details were checked against the research paper.