Astera Labs, Hyve Promise 89% More AI Tokens From the Same GPUs
Hyve Solutions, a TD SYNNEX subsidiary, will build Astera Labs' Orion compute boards, fabric trays and link boards into its air-cooled Apex racks, due in the first quarter of 2027. Astera says its single-switch fabric delivered up to 89% more tokens per second in an 8-GPU test, a potential cut to inference costs.
Astera Labs (ALAB) and Hyve Solutions, a subsidiary of TD SYNNEX (SNX), expanded their partnership on Monday to speed up rack-scale AI inference deployments. Hyve will build Astera's Orion connectivity hardware into its Apex server platform, and the companies say Astera's single-switch fabric delivered up to 89% more tokens per second than a conventional setup in testing, which could cut the cost of serving AI models.
The 89% claim
- The setup. Astera compared two 8-GPU systems that differed only in their fabric. One used a single Scorpio X-Series 320-lane switch on Astera's Orion Compute Board; the baseline used two conventional 144-lane switches.
- Why it helps. With one switch, data between GPUs takes a single hop instead of two, so the same GPUs spend less time waiting and produce more output.
- The fine print. Testing covered four large language models, three workload profiles and 1 to 512 concurrent requests. Gains of up to 89% varied by model and workload, and the companies did not name the GPUs used.
What Hyve is building
- The components. Hyve will integrate three Orion parts into Apex: the Orion Compute Board, which puts eight accelerators under one switch; the Fabric Tray, which shares the fabric across accelerators and pooled memory; and the Link Board, which extends the fabric from tray to rack.
- Faster rollouts. Because the parts come pre-validated, the companies expect shorter integration and qualification cycles for customers.
- Swap, don't rebuild. Buyers can pick their own CPUs and accelerators based on cost, lead time and workload, and upgrade accelerators later without rebuilding the system.
- No liquid cooling needed. Apex systems are air-cooled, so they can go into existing data centers. Hyve's Apex MGX 6U chassis is built to take higher-power accelerators as they arrive.
Built to scale
Orion, which Astera unveiled at the Open Compute Project summit in October, is a reference design rather than a product Astera manufactures itself; it reaches customers through server-building partners such as Hyve. Its 320-lane fabric is designed to scale from 8 to 160 accelerators per pod and supports flexible CPU and GPU mixes and memory pooling. The target buyers are neoclouds and enterprises that want inference capacity online quickly without committing to one chip vendor.
What the companies said
- Astera Labs. Pravin Patel, who leads Astera's AI Platform Solutions Group, said GPUs are too often left waiting on the fabric, and that customers can work with Hyve now to lower their cost per token.
- Hyve. CTO Dimitris Ziakas said “the rack is now the unit of AI,” and that pairing Apex with Orion lets customers deploy once and swap accelerators as workloads change.
What to watch
- Q1 2027 launch. Whether Orion-equipped Apex systems ship on time.
- Real-world results. Whether customer deployments match the up-to-89% throughput gain seen in testing.
- More partners. Which other server makers adopt Orion, and which neoclouds sign on.
- Competition. How open PCIe-based racks fare against single-vendor systems built around Nvidia's NVLink.
This article is informational only and is not investment advice.
Readers also read
Marvell, Penguin Bet on Recycled DDR4 and Light to Ease AI's Memory Crunch
AI InfrastructureTensorWave Lands 457 MW OpenAI Deal, First Neocloud in AMD's 6 GW Pact
Analyst CallsWall Street Cuts AI Chip Targets but Still Bets on Inference and Cybersecurity
Earnings PreviewNvidia Faces First Earnings Without Jensen Huang as Kress Takes the Call
MarketCatalyst is not a registered investment advisor and does not manage client assets. Content on this platform is provided for informational and educational purposes only. It is not investment advice, and MarketCatalyst is not a stock-picking or trade-alert service. Trading stocks and options involves risk, including the possible loss of principal. Consider your own goals, time horizon, and risk tolerance, and consult a qualified financial advisor before making any investment decision.
