SK hynix reported up to 5x faster time-to-first-token with its MARS technology at the Open Compute Project (OCP) event, alongside a capacity expansion of 14x for HBF memory. The numbers target the two bottlenecks that dominate AI inference economics today: latency at the start of a response and total memory capacity per module.

Time-to-first-token (TTFT) measures how long a model takes to produce its first output token after receiving a prompt. Users perceive it as responsiveness. Operators perceive it as cost, because every millisecond of stall ties up expensive accelerator time. A 5x improvement in this metric, if it holds in production, changes how quickly a fleet can turn over inference requests.

What is MARS and how does it work?

SK hynix presented MARS in the context of the OCP community's work on memory for AI systems. The company positions the technology as a way to cut the delay before a model begins generating output. The mechanism targets the memory path between storage and the compute units that run the model.

The claim is concrete: up to 5x faster time-to-first-token. The phrase "up to" matters here. It signals a best-case figure, likely under specific test conditions, rather than a guaranteed improvement across all workloads.

What does the 14x HBF capacity increase mean?

HBF, the memory format SK hynix discussed at the same event, expands capacity by 14x according to the company's presentation. Capacity per module sets how large a model a system can host and how many requests it can serve in parallel.

For operators, the practical question is simple: what does this change for deployment planning? A 14x jump in memory capacity means fewer modules per model, or more models per server, or larger models on existing hardware. It relaxes the tightest constraint in most inference clusters today.

Why the OCP stage matters

SK hynix chose the Open Compute Project venue for both announcements. OCP gathers hyperscalers and hardware engineers who define the specs for next-generation data center equipment. Presenting there signals that the company wants these technologies adopted as open infrastructure standards, not sold as closed products.

The announcements cover two different problem domains:

• MARS: latency reduction, measured as up to 5x faster time-to-first-token
• HBF: density, measured as 14x greater capacity

Together they address speed and scale, the two axes along which AI memory demand grows.

What comes next

The figures come from the manufacturer's own presentation. Independent benchmarks, customer deployments, and production data will determine whether the 5x and 14x numbers translate into real-world gains. Details on availability, pricing, and qualification timelines did not form part of the announcement as reported.

For now, the claims mark SK hynix's intent to stay central in AI memory as inference workloads, not just training, drive hardware decisions. The company competes in a market where memory bandwidth and capacity set the ceiling on model performance. These two numbers, 5x and 14x, are its latest bid to raise that ceiling.