Hardware · Taalas HC1 · Pre-COD
Inference cast in silicon
Phase 2+ watchlist at AGICY Vasilikos — model-specific ASIC optionality, not endorsement
Taalas hardwires a chosen model into custom silicon instead of streaming weights from HBM. The public HC1 technology demonstrator runs Llama 3.1 8Bon a TSMC 6 nm die (~815 mm², ~53B transistors) in a ~2.5 kW server class. Phase 1 CapEx remains Tenstorrent Galaxy. AMD announced a definitive agreement to acquire Taalas (6 Aug 2026) — subject to closing; AGICY treats HC1 as watchlist, not ordered fleet.
- Model cast into silicon
- Llama 3.1 8B demonstrator
- TSMC 6nm · 53B transistors
- Phase 2+ · on Watchlist
- No Taalas / AMD endorsement
- Vendor-claimed tok/s under diligence
- Not Phase 1 capital plan
- AMD acquisition agreement · not closed CapEx
Elevation · HC1 demonstrator · vendor photography · Phase 2+ watchlistTrade flexibility for tokens per joule
General-purpose XPUs move weights. Taalas casts a known model into the die — then serves that model at extreme per-user throughput.
HC1 is a technology demonstrator, not AGICY Phase 1 CapEx. It proves model-specific silicon can beat general GPUs on a fixed Llama 3.1 8B path — at the cost of re-spinning silicon when the model changes.VENDOR-CLAIMED
Founders include Tenstorrent alumni. That heritage is interesting diligence context — it does not make Taalas part of the Galaxy fleet plan.VERIFIED
Next: Architecture ledger →
Hardcore model silicon
Public product facts from Taalas — treated as vendor statements until AGICY lab validation exists.
- Process / die— TSMC 6 nm · ~815 mm² · ~53B transistorsVENDOR-CLAIMED
- Workload — Llama 3.1 8B hardwired (not arbitrary model swap)VENDOR-CLAIMED
- System— ~2.5 kW server class; no HBM / liquid-cooling claims for this demo SKUVENDOR-CLAIMED
- Roadmap posture — successors for larger models are marketing intent, not AGICY PODsDESIGN TARGET
Instantaneous inference — vendor ledger
Taalas publishes ~17k tokens/sec/user on Llama 3.1 8B for HC1, with GPU baselines measured by Taalas and peer LPU/wafer numbers from Artificial Analysis. Input sequence length cited: 1k/1k.
| Claim class | Status at AGICY | Notes |
|---|---|---|
| ~17k tok/s / user (8B) | Vendor-claimed | HC1 demonstrator — not installed AGICY capacity |
| Model flexibility | Hardwired to Llama 3.1 8B | Re-spin required for new models — different job than Galaxy |
| Phase 1 CapEx / P&L | Galaxy capital plan only | Taalas does not underwrite Round 1 fleet math |
HC1 vs Galaxy — different problems
Galaxy is the Phase 1 air-cooled RISC-V fleet. HC1 is a model-specific inference demonstrator under watchlist.VERIFIED
Taalas HC1
- Hardwired model silicon — extreme tok/s on one fixed model
- Technology demonstrator · Phase 2+ watchlist
- AMD acquisition agreement announced — roadmap may fold into Instinct / Helios systems
- Not Phase 1 capital / P&L baseline
Tenstorrent Galaxy
- General RISC-V inference servers — Phase 1 CapEx anchor
- Air-cooled · publicly priced · compiler path (TT-Metal)
- Model-flexible within the software stack
- Underwrites Round 1 campus fleet math
AMD definitive agreement — watch the close
On 6 Aug 2026 AMD announced a definitive agreement to acquire Taalas to strengthen specialized inference alongside Instinct GPUs and Helios rack-scale systems. Closing is subject to customary conditions.
- Primary source: AMD newsroom ↗
- Product page: taalas.com/products ↗
- Related AGICY eval: AMD Helios hardware page →
Next: AGICY relevance →
Where HC1 sits in the Vasilikos stack
Watchlist silicon for model-locked inference bursts — never a substitute for the Galaxy fleet plan.
- Phase 1— Tenstorrent Galaxy only for CapEx / P&L underwriting
- Phase 2+ — Taalas HC1 class evaluated beside Helios, Cerebras, d-Matrix, Trainium
- Honesty — pre-construction campus · vendor-claimed tok/s · AMD deal not closed CapEx