Hardware · AMD Helios · Pre-COD
Phase 2 evaluation at AGICY Vasilikos — silicon optionality, not endorsement
AGICY's technology map labels this lane AMD Helios · MI450. Publicly, AMD positions Helios as a rack-scale reference design: Instinct GPUs (MI455X product naming), EPYC host CPUs, Pensando networking, and OCP Open Rack Wide — not a discrete chip SKU, and not "Helios Semiconductor." Phase 1 capital planning remains Tenstorrent Galaxy; Helios sits on the Phase 2 evaluation queue under diligence.
Elevation · Helios ORW rack · Phase 2 evalHelios is a co-designed rack — not a drop-in Galaxy substitute and not a wafer-scale engine.
AMD published Helios as an open rack-scale AI infrastructure reference design on OCP Open Rack Wide (ORW), combining Instinct accelerators, EPYC CPUs, Pensando networking, and ROCm software. Volume deployments are publicly targeted for 2H 2026.VENDOR-CLAIMED
AGICY keeps Helios on the map because open rack standards and dense HBM training/inference are a different facility and software problem than air-cooled RISC-V Galaxy floors — not because AMD "wins" Phase 1 underwriting.VERIFIED
Next: Architecture stack →
One scale-up domain: GPUs, CPUs, and networking co-designed into an ORW liquid-cooled rack.
Next: Memory & fabric →
Dense HBM4 and rack-scale fabric are the Helios bet — under independent diligence before any CapEx.
AMD materials cite order-of-~31 TB HBM4 and hundreds of TB/s scale-up bandwidth across the 72-GPU domain. Treat these as vendor-claimed until AGICY technology committee validation.VENDOR-CLAIMED
ROI / planner price bands (if any) live in illustrative YAML only — they are not live capital baseline and are not commitments on this page.VERIFIED
Next: Best use cases →
Honest AGICY buyer framing — what public Helios specs support, tagged by workload. Not live campus SLAs.
AMD positions Helios as a 72-GPU scale-up domain for frontier model training (ROCm + PyTorch / JAX). Strong Phase 2 evaluation fit when buyers need GPU rack-scale train pods — not AGICY Phase 1 CapEx, and not a live training SLA.
Public AMD materials list fine-tuning alongside training and inference. Dense HBM4 and a single-hop fabric suit large full fine-tunes or continued pre-train jobs that outgrow multi-node Ethernet clusters — still under diligence, ROCm maturity dependent.
Vendor thesis: ~31 TB HBM4 / rack for larger models, long contexts, and multi-agent workflows; tokens-per-dollar claims vs NVIDIA Vera Rubin NVL72 are AMD-modeled, not independent production benches. Complementary to Galaxy civil decode, not a substitute.
Same ORW liquid rack can host train jobs by day and high-batch inference by night — useful Phase 2 campus optionality beside Galaxy Phase 1 floors. Pre-COD plan only; no claim AGICY already runs Helios.
OCP Open Rack Wide, UALink / UALoE, Ultra Ethernet — open rack path vs closed NVIDIA stacks. Trade-off: CUDA ecosystem maturity and tooling still favor NVIDIA for many teams; treat ROCm porting as a selection-gate cost.
Pensando scale-out bandwidth is vendor-claimed for jobs that span racks (train or distributed infer). Credible on paper for AGICY Phase 2 facility studies; power, cooling, and OEM pack-out still VERIFY before any CapEx commit.
Strengths vs NVIDIA training stacks (vendor / analyst framing): denser HBM and open ORW / UALink path; claimed scale-out headroom. Weaknesses: CUDA ecosystem maturity, ROCm port cost, and performance claims that are largely AMD-modeled vs Vera Rubin NVL72 — not third-party production benches. Vs Cerebras: different architecture (GPU rack-scale vs wafer-scale) — both Phase 2 eval, neither Phase 1 CapEx.VENDOR-CLAIMED
Helios = Phase 2 GPU rack-scale training-eval lane. Galaxy = Phase 1 decode/inference CapEx anchor. They do not underwrite the same case.VERIFIED
Also on the multi-silicon queue: Cerebras CS-3 (wafer-scale train/infer eval), IBM z17 (AI next to the ledger), and d-Matrix Corsair (decode-lane DIMC).
Liquid ORW racks belong in Phase 2/3 expansion halls — not Phase 1 air-cooled Galaxy floors.
Helios-class density implies liquid cooling, wider ORW service envelopes, and technology-committee gates before CapEx. Mediterranean free-cooling and the campus power roadmap can support high-density pods under EU jurisdiction if selection criteria clear — still pre-COD.DESIGN TARGET
Sovereign AI Campus briefing →
Next: Diligence & honesty →
What AGICY claims — and what it does not — about AMD Helios on campus.
Multi-silicon sovereign posture (pre-COD) — open rack-scale optionality beside the Galaxy Phase 1 anchor.
Lease or finance Helios units on AGICY premises, or at your site if it meets our T&Cs. Free estimate for integration and delivery (quote-only rack-scale). Pre-COD; Leasing ≠ SRA / CY tokens.