AI Training: No Need to Buy More GPUs? The Secret to Finding 'Wasted Compute'

In the recent AI frenzy, the most highly sought-after commodity is undoubtedly the Graphics Processing Unit (GPU)—hardware that processes complex mathematical calculations at high speeds. Companies around the world are desperate to secure GPUs, spending astronomical amounts of money to train their AI models. It is much like the Gold Rush era, when people scrambled to acquire pickaxes to strike gold. But what if the GPUs you already possess are actually failing to deliver even half of their potential performance?

Expanse, the startup we are introducing today, was born from this very question. They have developed an “intelligence layer” (software that controls and manages infrastructure efficiency) that allows companies to drastically increase AI training efficiency using only the infrastructure they already own, without purchasing new hardware. Reference 1, Reference 5

Why Is This Important?

For businesses, AI training is a fierce battle of “time” and “cost.” The price per GPU is skyrocketing, and the operational costs of the infrastructure managing them are no small matter. What if you could boost the efficiency of your current resources by just 30% through Expanse? Reference 9 This has an economic impact equivalent to investing billions of dollars in new hardware. Reference 5

Furthermore, predictable performance is directly linked to service stability. AI-driven companies fear sudden training interruptions or system failures above all else, and Expanse helps prevent these risks by predicting potential hazards at the task submission stage. Reference 5

In Simple Terms

Let’s compare Expanse’s role to a massive restaurant kitchen. This kitchen employs dozens of top-tier chefs (GPUs). However, because the kitchen is so busy, nobody knows which order should be assigned to which chef for the fastest turnaround. Orders (AI training tasks) keep piling up; some chefs are idle while others are overloaded and struggling.

Expanse acts as the “veteran manager” of this kitchen. This manager monitors the condition of every chef in real-time, accurately grasping how long each dish will take to prepare and which chef is currently exhausted and at high risk of breaking down (failure risk). Reference 2, Reference 5 When an order comes in, the manager immediately issues instructions, saying, “This task should be assigned to this chef for maximum efficiency.” As a result, the cooking speed of the entire kitchen becomes significantly faster.

Technically, Expanse is installed on every computer in the data center, meticulously monitoring the real-time state of hardware (DCGM, CUPTI, etc.). It is similar to collecting various metrics displayed on a car’s dashboard to check its condition. Reference 2 Based on this data, it builds a “digital map” of how the current infrastructure is performing and finds the optimal path for the next task. Reference 6

Current Status

Expanse is a startup backed by Y Combinator (YC), Silicon Valley’s leading accelerator, and is currently gaining significant attention in the AI industry. Reference 2, Reference 7 They are already improving efficiency in actual high-performance computing (HPC) environments by integrating with data center standard schedulers such as SLURM or Kubernetes (a program that manages data center computer resources). Reference 2, Reference 5

For companies struggling with insufficient hardware, resource acquisition has become a strategic core—so much so that “GPU is the new oil”—and Expanse is teaching them how to use these precious resources without waste. Reference 3

What Lies Ahead?

AI training models will continue to grow larger and more complex. Consequently, efficient infrastructure management will become a matter of survival, not choice, for companies. As Expanse is applied to more large-scale clusters, it is expected to spread a “software-centric” mindset, where companies optimize infrastructure more smartly rather than simply buying hardware at a faster rate. Perhaps it is thanks to such “veteran manager” solutions that the AI services we use can operate more affordably and stably. Reference 5

MindTickleBytes’ AI Reporter Perspective

Software technology that pushes hardware performance to its limits has always accelerated human technological progress. The emergence of Expanse is an interesting indicator that the AI industry has moved from a stage of “quantitative expansion” to one of “qualitative management.”

References

1. [Launch YC: Expanse - Unlock wasted GPU capacity

   Y Combinator](https://www.ycombinator.com/launches/QCF-expanse-unlock-wasted-gpu-capacity)

2. Launch HN: Expanse (YC P26) – Unlock Wasted GPU Capacity
3. Expanse · YC Spring 2026
4. progscrape: gpu

5. [Expanse

   Intelligence Layer for HPC and GPU Clusters](https://expanse.sh/)

6. Expanse is the intelligence layer for compute infrastructure that…

7. [Natural 20 — AI News in Real-Time

   The Bloomberg Terminal for AI](https://natural20.com/c/m6r0pc)

8. Запуск HN: Expanse (YC P26) – Раскройте неиспользуемые мощности GPU - TheNote.app

9. [30 % mehr GPU-Leistung: Wie Expanse HPC revolutioniert

   WAI News](https://wainews.com.br/posts/30-mehr-gpu-leistung-wie-expanse-hpc-revolutioniert)