OpenAI hardware chief Richard Ho’s message is that cheaper models do not necessarily mean less computing overall. At a Synopsys SNUG keynote, he said scaling laws would continue to increase compute needs as work shifts from training the largest models toward post-training and reasoning at inference time. That points to ongoing demand for accelerators and the infrastructure around them—but it does not establish a specific forecast for chip sales, data-center spending, or consumer prices.
What did Richard Ho mean by scaling laws continuing?
Scaling laws describe how model capabilities can improve as more resources are applied. Ho said, “It does appear that scaling laws will continue to grow [compute needs] to provide extra capabilities,” as quoted by EE Times. His point was not that every AI model must become larger. Rather, the compute burden can keep growing as companies use more computation after initial training, including to refine models and let them spend more time generating or checking answers.
That distinction matters: a smaller or less expensive model can lower the cost of one task while broader use, more demanding tasks, or longer reasoning workloads increase total demand. Efficiency gains and growth in aggregate computing needs can therefore happen at the same time.
Where is the computing demand shifting?
From frontier training toward post-training
Training a frontier model is only one part of its development. Ho described a shift toward post-training work: computation used to adapt and improve a model after its initial training. The keynote account does not quantify how much compute this takes relative to frontier training, so it supports a direction of change, not a precise split.
Recommended Free Tools
#1 Best Overall
More computation while answering
Test-time compute is the computation used when a model responds to a prompt. Reasoning-oriented systems may generate more tokens or perform additional steps before returning an answer. That can make a single response more compute-intensive, even if the underlying model is smaller or cheaper to run than an earlier alternative.
How fast has AI compute been growing?
EE Times reports estimates from Epoch AI that AI training compute grew 6.7 times per year through 2018 and more than four times per year after 2018. These are historical growth estimates attributed to Epoch AI by EE Times—not a forecast that the same rate will continue. The article connects growth to factors including Moore’s law, reduced-precision computation, larger systems, and the ability to run jobs for longer.
Rank #2
- Ideal for Gifting
- Ideal for a bookworm
- Compact for travelling
Those figures describe compute growth, not a guaranteed rate of growth in GPU revenue, electricity use, or any individual company’s capital spending. They also do not show how much future demand will be offset by more efficient models and hardware.
What does continued scaling imply for AI hardware?
GPUs remain important, but peak specifications are not the whole story
Ho described hardware development as full-stack co-design: the model, compiler, chip, system, and kernels must work together. A chip’s advertised peak performance does not necessarily translate into useful throughput if another part of the system is the bottleneck. Memory capacity and bandwidth, latency, networking, and software compatibility all affect what a deployed system can deliver.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Custom accelerators can be strategically useful, but they are not a standalone answer. Their value depends on how well the hardware fits the workloads and whether compilers, kernels, and systems can use it effectively.
Large systems need reliable networking and operations
Ho’s remarks, as reported by EE Times, describe warehouse-sized computers today and the prospect of larger infrastructure. Training can span clusters in different geographies, making communication between machines and dependable operation important. If a synchronous training job depends on many components, a failure in one part can stall the broader job; reliability and uptime therefore affect productive compute, not just convenience.
Rank #4
Chip development must keep pace with software
EE Times reports Ho’s observation that chip-design cycles of roughly 18–24 months are slow compared with AI research. That mismatch puts pressure on hardware teams to shorten the path from architecture to tape-out, while still designing for workloads that may evolve before a chip is ready.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What should a buyer or investor watch?
Ho’s comments describe technical pressures, not investment advice or a forecast for a particular company. To evaluate an AI hardware or infrastructure option, compare the dimensions that determine useful performance and cost:
Best Value
- It can be a gift option
- Comes with secure packaging
- Helpful in various ways
- Throughput and latency: how much work the system completes and how quickly it responds.
- Memory: capacity and bandwidth, which can constrain model size and data movement.
- Power efficiency: useful computation delivered for the power consumed.
- Software fit: compatibility with the compilers, kernels, and model workloads a user needs.
- Networking and scale: how well multiple machines communicate as jobs grow.
- Reliability: whether the system can sustain long, distributed workloads without costly interruptions.
- Total system cost: the cost of the complete deployment, not just the accelerator chip.
A consumer graphics card is not a direct proxy for an accelerator deployment at OpenAI scale. Datacenter systems also depend on memory, networking, power, cooling, software, and coordinated operation across many machines. Ho’s argument supports watching the broader infrastructure stack; it does not establish which vendor or investment will benefit most.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
