Cost of Compute 2026: H200, B200, MI325X, and the TPU v6 Trendline

What's Cheap and What's Not

Compute costs for AI workloads in 2026 are dropping fast for inference and roughly flat per-FLOP for training. The mix of available hardware has broadened: NVIDIA still dominates but AMD and Google have gained share. This piece walks through the 2026 numbers and where the curves are heading.

The Hardware Lineup

flowchart TB
    NV[NVIDIA] --> H100[H100<br/>2022-2024 mainstream]
    NV --> H200[H200<br/>2024-2026 mainstream]
    NV --> B200[Blackwell B200<br/>2025-2026 frontier]
    NV --> GB[GB200 NVL72<br/>rack-scale]
    AMD[AMD] --> MI300[MI300X<br/>2024]
    AMD --> MI325[MI325X<br/>2025]
    AMD --> MI355[MI355X<br/>2026]
    Goo[Google] --> TPU5[TPU v5p<br/>2024]
    Goo --> TPU6[TPU v6 'Trillium'<br/>2025-2026]

Per-FLOP Trends

For BF16/FP8 throughput per dollar, the rough 2026 picture (numbers approximate, vary by deal):

• H100: baseline (1.0x)
• H200: ~1.2x
• B200: ~2.5-3x H100 per dollar at FP8
• MI355X: ~2-2.5x H100 per dollar at FP8, FP4 native
• TPU v6: ~2x TPU v5p, comparable to mid-tier GPU economics

For FP4 training and inference (Blackwell native, MI355X native, TPU v6 partial), the per-FLOP cost is roughly 2-3x cheaper still.

Per-Token Inference Cost

Per-million-token inference cost for a 70B-class model in 2026:

• $0.10-0.40 cents range for hosted providers
• Self-hosted on Blackwell: comparable when amortized
• Open-source models on cheap GPU rentals (Lambda, RunPod): substantially cheaper for batch workloads

The 2026 inference cost curve has dropped roughly 5-10x from 2024 levels for comparable quality. Training costs have dropped less — perhaps 2x for like-for-like compute.

Memory Is the Constraint

flowchart LR
    Param[Parameters] --> Mem[Memory required]
    Mem -->|HBM3e is fast and expensive| Cost[Cost dominated by memory]

The dominant cost in 2026 inference is memory bandwidth, not raw compute. HBM3e capacity per GPU varies:

• H100: 80GB HBM3
• H200: 141GB HBM3e
• B200: 192GB HBM3e
• MI355X: 288GB HBM3e
• TPU v6: 32GB per chip but liquid-cooled clusters scale memory across many chips

Larger memory per GPU lets you fit larger models on fewer cards, dropping serving cost.

The Power Wall

By 2026, data center power is the binding constraint for new AI capacity in many regions. A B200 rack runs at 120-140 kW; existing data centers were not built for this density. The capex shift toward purpose-built AI campuses (Microsoft, Meta, Amazon, OpenAI, x.AI) is partly a response.

This is a topic in itself; covered in the next article.

Choosing Hardware in 2026

flowchart TD
    Q1{New deployment?} -->|Yes| Q2
    Q1 -->|No, existing H100 fleet| Keep[Keep until H100 depreciated]
    Q2{Frontier training?} -->|Yes| BTPU[B200 or TPU v6]
    Q2 -->|No, inference| Q3{Cost optimized?}
    Q3 -->|Yes| MI[MI355X or hosted]
    Q3 -->|No, lowest latency| B[B200]

What's Coming

• B300 / Rubin (NVIDIA): expected late 2026 / 2027
• MI400 series (AMD): expected late 2026
• TPU v7: expected 2027
• More aggressive FP4 and beyond (FP3? Sub-bit-precision?) on next-generation hardware

The doubling cadence of AI compute capacity per dollar that drove 2022-2025 is showing signs of slowing as we approach physical limits. 2026-2028 will be a slower curve than 2022-2025.

What This Means for Builders

For most teams, the action is straightforward:

• Use the best inference hardware your provider offers (B200, MI355X, TPU v6)
• Quantize to FP4 / FP8 wherever quality allows
• Cache aggressively (prompt caching, KV-cache, response caching)
• Monitor per-task cost; the trend is your friend

For teams that own infrastructure: the 2024 H100s are still useful but the depreciation schedule should reflect that B200 / MI355X are 2-3x cheaper per equivalent throughput. New deployments should default to current generation.

Sources

• NVIDIA GTC announcements — https://www.nvidia.com/gtc
• AMD Instinct roadmap — https://www.amd.com/en/products/instinct
• Google TPU v6 — https://cloud.google.com/tpu
• "AI compute trends" Epoch AI — https://epochai.org
• "Artificial Analysis" benchmarks — https://artificialanalysis.ai

Cost of Compute 2026: H200, B200, MI325X, and the TPU v6 Trendline: production view

Cost of Compute 2026: H200, B200, MI325X, and the TPU v6 Trendline is also a cost-per-conversation problem hiding in plain sight. Once you instrument tokens-in, tokens-out, tool calls, ASR seconds, and TTS seconds against booked-revenue per call, the right tradeoff between Realtime API and an async ASR + LLM + TTS pipeline becomes obvious — and it's almost never the same answer for healthcare as it is for salons.

Broader technology framing

The protocol layer determines what's possible: WebRTC for browser-side widgets, SIP trunks (Twilio, Telnyx) for PSTN voice, WebSockets for the Realtime API streaming session. Each has its own jitter buffer, its own ICE/STUN dance, and its own failure modes when a customer's corporate firewall is hostile.

Front-end is Next.js 15 + React 19 for the marketing surface and the in-app dashboards, with server components used heavily for the SEO-critical pages. Backend splits across FastAPI for the AI worker, NestJS + Prisma for the customer-facing API, and a thin Go gateway that does auth, rate limiting, and routing — letting each service scale on its own characteristics.

Datastores: Postgres as the source of truth (per-vertical schemas like healthcare_voice, realestate_voice), ChromaDB for RAG over support docs, Redis for ephemeral session state. Postgres RLS enforces tenant isolation at the row level so a misconfigured query can't leak across customers.

FAQ

What's the right way to scope the proof-of-concept? Setup runs 24 hours, the trial is 7 days with no credit card, and pricing tiers are $49, $99, and $149 — so a vertical-specific pilot is a same-week decision, not a quarterly project. For a topic like "Cost of Compute 2026: H200, B200, MI325X, and the TPU v6 Trendline", that means you're not starting from scratch — you're configuring an agent template that's already been hardened across thousands of conversations.

How do you handle compliance and data isolation? Day one is integration mapping (scheduler, CRM, messaging) and prompt tuning against your top 20 real call transcripts. Day two through five is shadow-mode running, where the agent transcribes and recommends but a human still answers, so you can compare side-by-side. Go-live is the moment your eval pass-rate clears your internal bar.

When does it make sense to switch from a managed model to a self-hosted one? The honest answer: it scales until your tool catalog gets stale. The agent is only as good as the integrations it can actually call, so the operational discipline is keeping schemas, webhooks, and fallback paths green. The platform handles the rest — observability, retries, multi-region routing — without your team owning the GPU layer.

Talk to us

Explore a live demo and compare current plans to find the right fit for your business.