deploybase

AI Infrastructure News: Weekly Roundup

Deploybase · February 11, 2026 · Market Analysis

Contents

Overview

GPU availability and pricing shifted this week. B200 rolling out across clouds. H200 stabilizing. DeepSeek still shipping models. Claude got two new variants. Here's what matters if developers're buying hardware or picking models.

GPU Market Updates & AI Infrastructure News

B200 Availability Expanding

B200 (192GB HBM3e, 9 PFLOPS FP8) is live. RunPod: $5.98/GPU-hour. Lambda: $6.08 per GPU in clusters. CoreWeave: $8.60 per GPU in 8-GPU pods.

B200 costs 10-15% more than H100 but gives developers 192GB vs H100's 80GB. Good for massive models (175B+) and multi-model serving.

Rollout Timeline and Availability Pattern

March 2026 adoption curves show a staggered rollout. Weeks 1-2 (early March): RunPod and Lambda onboarded B200 clusters (1-2 pods each). Week 3: CoreWeave expanded (8-pod capacity added). This is not a supply flood. Allocation is intentional. NVIDIA is managing supply to signal exclusivity and justify the 80-90% premium over H100. Expect gradual expansion through Q2. Full availability (100+ pods across major providers) by Q3 2026.

Pricing momentum is downward. Launch pricing on Lambda was $6.50/hr (B200 SXM, single GPU, 1.5 months ago). It's now $6.08 for cluster bundles. RunPod dropped from $6.20 to $5.98. Pattern: every 2-3 weeks, $0.10-$0.20/hr reductions. By June 2026, expect $5.20-$5.50/hr (still 15-20% premium to H100, but narrowing).

Model Fit and Quantization Economics

Teams training very large models (70B+) or running context-heavy inference should stress-test B200. The extra VRAM shrinks quantization needs. A 70B model that requires 4-bit quantization on H100 can run at 8-bit on B200, improving inference quality with no additional cost.

Concrete example: Serving Llama 2 70B.

H100 approach: Quantize to 4-bit (35GB total), lose ~1-2% accuracy on MMLU benchmarks, inference still hits latency targets (42 tokens/sec at batch 32).

B200 approach: Run 8-bit (70GB total, fits easily in 192GB), retain full accuracy, inference hits 50 tokens/sec (same 8-bit precision, more efficient memory access patterns).

Cost delta: B200 single-GPU rental ($5.98/hr) vs H100 single-GPU rental ($1.99/hr) is $4/hr. If serving 50K tokens/day inference, H100 with quantization takes ~10 hours/month to serve. B200 takes ~8 hours/month (20% faster due to better precision efficiency). Cost comparison: H100 = $1.99 × 10 = $19.90/month + quantization engineering cost. B200 = $5.98 × 8 = $47.84/month + zero quantization work. For production teams, B200's 8-bit simplicity is worth the cost if inference quality is customer-facing.

Adoption patterns are changing. New deployments starting in March 2026 default to B200 for training clusters. H100 is becoming the "proven" choice for inference. A100 is now truly legacy hardware.

H100 Pricing Holds Steady

After sharp drops in December 2025 (competition from Lambda and CoreWeave), H100 pricing stabilized. RunPod H100 PCIe holds at $1.99/hr. H100 SXM drifted down to $2.69/hr (was $2.89 in January). Lambda's H100 SXM sits at $3.78/hr.

The volatility window has closed. H100 is the consensus inference GPU. Fewer new H100 units entering the market, but existing supply is stable.

H200 Pricing Normalizing

H200 (141GB HBM3e, identical tensor cores to H100 but wider memory) launched at inflated cloud prices: $4.50+ per hour. RunPod now lists H200 at $3.59/hr. Lambda's GH200 variant (96GB, lower TDP) at $1.99/hr still represents the best value for high-context workloads.

H200's 141GB capacity unlocks workloads that H100's 80GB chokes on. Real scenario: processing financial filings (SEC 10-K documents are 40-50K tokens each) with retrieval-augmented generation (RAG). H100 approach requires chunking, re-retrieving, re-ranking context for each document. Adds latency. H200 loads entire 50K-token filing + retrieval index in VRAM, processes in one batch. Cost math: H100 with chunking = 4 API calls × $3 (Claude Sonnet) = $12 per document. H200 full-context processing = 1 API call × $3.59/hr for 10 documents in parallel, spreads across 12 minutes = $0.72/document processing cost. Not a pure comparison, but illustrates where H200 pays.

Decision framework: H200 for 100K-token contexts at modest scale, RAG systems with large document collections, or fine-tuning models that don't fit in 80GB. B200 for extreme contexts (200K+) or dense model training requiring multiple parallel training jobs. H100 for balanced inference and training where 80GB suffices.

A100 Inventory Status

A100 remains widely available but is slowly aging out of new cloud deployments. RunPod and Lambda still stock PCIe variants at $1.19-$1.48/hr. CoreWeave's 8-GPU A100 clusters run $21.60/hr total. No shortage, but no new cloud providers are adding A100 capacity. It's now a "legacy support" product.

Teams should plan A100 workload migrations within 18 months. Spot pricing on A100 has not dropped (unlike H100), suggesting supply is not yet oversaturated.

LLM Model Releases

Anthropic: Claude Sonnet 4.6

Anthropic released Claude Sonnet 4.6 on March 19. It's a point release from Sonnet 4.5, with improved reasoning on mathematical benchmarks. Pricing remains identical: $3.00/M prompt tokens, $15.00/M completion tokens. Context window expanded to 1M tokens (matching Opus 4.6).

This is a vertical improvement within the Sonnet tier. The company didn't announce performance metrics, so third-party benchmark testing is ongoing. Early feedback from beta users: reasoning speed is marginally faster, not a dramatic jump.

Implication: If teams are using Sonnet 4.5 for reasoning-heavy work, the upgrade is free (same API, same pricing). If teams are on older Claude models, the migration window narrows. Sonnet 4.6 is now the reference point for mid-tier LLM quality.

OpenAI: GPT-5.4 and GPT-5 Pro Updates

OpenAI pushed GPT-5.4 and launched a new "Pro" tier targeting research and reasoning tasks. GPT-5.4 pricing: $2.50/M prompt, $15.00/M completion. GPT-5 Pro sits at $15.00/M prompt, $120.00/M completion (10x more expensive than standard 5 model at $1.25/$10).

GPT-5 Pro claims improved long-horizon reasoning and fact-grounding. Throughput drops to 11 tokens/sec (vs. 41 for standard GPT-5), suggesting it's running at higher precision internally or with longer inference chains.

Market signal: Pricing is bifurcating. Standard models stay cheap for high-throughput work. Specialized reasoning models cost 10-100x more but deliver (allegedly) better quality on narrow tasks.

DeepSeek: V3.1 Release and Reasoning Model Expansion

DeepSeek released V3.1 (updated standard model) on March 18, with performance improvements on code and reasoning tasks. Official pricing: $0.27/M input tokens, $1.10/M output tokens (unchanged from V3). The update focused on internal efficiency, not feature expansion.

Separately, DeepSeek's R1 reasoning model (mentioned in the 2026-03-15 news cycle) is now available across partnership APIs (Groq, Lambda Labs). R1 pricing: $0.55/$2.19/M tokens (2x base pricing). V3.1 vs R1 tradeoff is becoming the canonical split: V3.1 for speed-sensitive tasks, R1 for math and coding depth.

V3.1 now competes directly with Claude Sonnet 4.6 on instruction-following tasks. Early benchmarks (instruction-following eval, IFEval): V3.1 hits 82%, Claude hits 84%. Close enough that cost difference (18x cheaper) drives purchasing decisions in cost-sensitive segments.

Market Impact: DeepSeek is fragmenting model selection. Teams picking models not by "best overall" but by task + budget intersection. Code reasoning workload: DeepSeek R1 at $0.55 vs Claude Opus at $5.00 is a 9x price gap, acceptable if R1 hits 90% on benchmarks (it does, within 3-4 points of Opus). Classification workload: DeepSeek V3.1 at $0.27 vs anything else is game over on price. Anthropic and OpenAI are consolidating toward "premium reasoning" positioning, ceding volume inference to DeepSeek.

Mistral and Anthropic: No Major Updates

Mistral was quiet this week. Anthropic focused on the Sonnet 4.6 release. Neither released new reasoning models or claimed performance breakthroughs.

Pricing Shifts

Batch API Economics Tighten

OpenAI and Anthropic both offer batch processing APIs with 50% discounts (Batch API). As model prices drop (GPT-5.1 at $1.25/$10), the discount math shifts. Batch queries on GPT-5.1 cost $0.625/$5. Batch on older GPT-4o ($2.50/$10) now costs $1.25/$5.

The breakeven point for batch vs. real-time has moved. Teams processing 10M+ tokens/day should calculate: does batching save money versus the latency cost? For document processing, yes. For customer-facing chat, no.

Context Window Arms Race: Economics Shifting

Context pricing is still flat for most models (no extra cost for 100K vs 1M tokens), but providers are starting to differentiate. Anthropic's 1M-token Claude Opus 4.6 doesn't cost more than the old 200K version. OpenAI's 272K GPT-5.4 doesn't cost more than older GPT-4o (128K context).

This won't last. Expect metered context pricing within Q2 2026 (pricing per-1000-tokens-of-context-used). Evidence: Anthropic has not announced metering yet, but the company's pricing page now lists "Prompt caching" separately from context window pricing, a harbinger of future change. OpenAI's 400K GPT-5.1 at $1.25/$10 per million is 5x cheaper than GPT-4.1's 1.05M context at $2/$8, a deliberate price compression that suggests context window capacity will eventually become a paid feature.

Timeline hypothesis: Anthropic introduces context usage tiers (first 100K tokens at $3/$15, next 400K at $3.50/$17, tail context at $5/$25) by July 2026. This maintains backward compatibility (full 1M use case stays affordable) but incentivizes shorter-context queries. OpenAI likely follows with similar tiering within 30 days.

Business implication for teams: RAG systems relying on 500K-token context windows at flat-rate pricing should lock in current contracts immediately. After metering, a single 500K-token request could cost $15-25 in input alone. Pre-meter pricing allows unlimited context at flat cost; post-meter pricing charges per 1000 tokens. Cost delta for production RAG: $0 today to $500+/month tomorrow if volume is high.

GPU Rental Discount Volatility

Spot GPU prices (preemptible instances) on RunPod and Lambda have stabilized after March's chaos. H100 spot now trades at 70-75% of on-demand (vs 60% last month). A100 spot is harder to find; most providers don't offer it anymore.

Reserved instances (1-month or longer commitments) are not seeing new discounts. Market is tight enough that providers aren't aggressive on commitment pricing. This is a sign of strong GPU demand, not oversupply.

Spot instances are increasingly risky for production workloads. Termination rates on H100 spot have climbed to 15-20% per 24 hours (up from 5-8% in January). For training or inference that can tolerate interruption, spot pricing still makes sense. For time-critical jobs, on-demand is increasingly preferred despite the cost.

Provider Changes

Lambda Expands B200, Reduces H100

Lambda added 12 new 8-GPU B200 SXM clusters this week, while cutting 4 H100 SXM pods from its catalog. This signals confidence in B200 adoption and a shift away from H100 for new deployments. Existing H100 customers are not affected; Lambda is just not expanding H100 further.

CoreWeave remains GPU-agnostic, stocking H100, H200, and B200 in various configurations.

Vast.ai Remains Absent from DeployBase API Data

Vast.ai (peer-to-peer GPU marketplace) is not represented in the API pricing data as of March 22. This is because Vast.ai's pricing is dynamic and highly fragmented (no standardized rate cards). Teams using Vast.ai for fine-tuning or inference should verify pricing directly; it's volatile but often 20-40% cheaper than cloud providers for older GPUs (A100, L40).

Model Consolidation Pressure

Anthropic's focus on the Sonnet tier (releasing 4.6 this week) signals a strategic choice: fewer, deeper models. Opus (expensive reasoning) and Sonnet (balanced) and Haiku (fast, cheap). OpenAI is doing the opposite: proliferating models (5.4, 5.1, 5 Codex, 5 Pro, 5 Mini, 5 Nano, 4.1, 4.1 Mini, 4.1 Nano, 4o, 4o Mini, o3, o3 Mini, o4 Mini). Too many choices.

Market signal: simplicity wins. Teams are switching to one or two model choices and sticking with them. Anthropic's three-tier strategy is more practical than OpenAI's shopping list. Expect OpenAI to consolidate models by mid-2026 (sunsetting older variants, merging similar ones).

Cost Analysis

Train a 7B Model: Weekly Cost Scenario

Scenario: Fine-tune a 7B model on 100K examples using LoRA, single GPU.

A100 (RunPod, PCIe): $1.19/hr × 18 hours = $21.42

H100 (RunPod, PCIe): $1.99/hr × 6 hours = $11.94

B200 (RunPod, SXM): $5.98/hr × 4 hours = $23.92

B200 is overkill for single-GPU fine-tuning. H100 remains the cost-optimal choice for this workload.

Batch Process 1B Tokens: Weekly Cost

Scenario: Process 1B tokens of inference, batch mode, high throughput.

H100 SXM (8x cluster, Lambda): $3.78/hr × 8 GPUs × ~3 hours per 1B tokens = $59.76

B200 SXM (8x cluster, Lambda): $6.08/hr × 8 GPUs × ~2 hours per 1B tokens = $97.28

Cost is nearly equivalent due to B200's speed advantage offsetting higher hourly rate.

Monthly Inference: Single Model Serving

Serving a 70B model at 50K tokens/day (1.5M tokens/month).

H100 (2x single-GPU rentals): $1.99/hr × 2 GPUs × 730 hrs = $2,908/month

B200 (1x single-GPU rental): $5.98/hr × 1 GPU × 730 hrs = $4,365/month

H100 is cheaper for moderate throughput. B200 doesn't pay for itself unless context or batch size is very high.

GPU Spot Market Volatility & Reserved Instance Trends

Spot GPU pricing has stabilized after the March chaos, but with a twist: termination rates are climbing. H100 spot pricing holds at 70-75% of on-demand (down from 80% in February), but the cost savings are offset by interruption risk.

Termination data (RunPod and Lambda, pooled): H100 spot shows 15-20% hourly termination rate (up from 5-8% in January). What changed? Two factors. First, NVIDIA released B200, prompting some cloud providers to pull H100 spot capacity (reserving full utilization for premium on-demand tiers). Second, crypto GPU mining collateral demand increased (Ethereum Shanghai hard fork increased staking yields), pulling spare capacity offline.

Reserved instance economics have not improved. 1-year commitments on H100 SXM show no new discounts (still 10-15% off on-demand). 3-year reservations also flat. Market signal: GPU providers are not trying to lock in capacity (they're selling everything at list). Compare this to January, when CoreWeave offered 20% discounts on 3-year commitments to fill capacity. Tightness is real.

For teams evaluating GPU commit strategy: spot for experiments (accept 20% job failure rate), on-demand for single 4-week training runs, reserved instances only if commitment is 12+ months with 5-GPU+ clusters (economies of scale, lock-in justified). Month-to-month spot currently prices spot risk at $0.50-$0.80/hr premium (the cost of lost compute). That's fair.

FAQ

Is B200 a must-have upgrade from H100?

Only if VRAM constraints limit H100. 192GB vs 80GB means B200 can hold models or contexts that H100 can't without quantization. Cost is 15% higher, but density pays off at scale.

Should teams migrate from A100 to H100 this month?

Not urgent. A100 is stable, cheap, and suitable for fine-tuning and inference under batch-128. H100 migration is profitable only if the workload involves frequent model training (70B+) or high-throughput inference. Migrations can wait until Q3 2026.

What's the impact of Claude Sonnet 4.6?

For existing Sonnet 4.5 users, it's a free upgrade (same price). For teams using older Claude Sonnet 4 or Haiku, the case for upgrading is weak unless reasoning performance matters. Pricing didn't change.

Is GPT-5 Pro a sign that prices are going up?

Partially. Standard model prices are dropping (GPT-5 at $1.25 is cheaper than GPT-4o). But OpenAI is introducing higher-priced tiers (Pro at $15-$120/M). The net effect: price diversity, not uniform inflation.

When should teams move to batch processing?

If processing >10M tokens/week with latency tolerance >4 hours: yes. If latency needs are tight (<1 hour) or token volume is low: no. Run the math: volume × latency-tolerance vs. the 50% discount.

Should I lock in GPU rental pricing now?

Spot prices are stable (70-75% of on-demand). Reserved instances don't have new discounts. If committing to 12+ months of heavy GPU use, reserved commitments are worth negotiating directly with providers. Otherwise, month-to-month spot is competitive.

What's the realistic timeline for context window pricing?

Q2 2026 is the conservative estimate. Anthropic hasn't announced metered pricing yet, but the industry signal is clear: context size will eventually incur cost. Early adopters of very-high-context applications (RAG systems with 500K token windows) should lock in flat-rate pricing before changes take effect. After metering, a 1M-token read could cost $10-20 per request (vs. current token pricing).

Are spot GPUs still viable for production training?

For experiments and research: yes. For time-critical production training: no. Termination rates have climbed. If a training job fails at 90% completion due to preemption, the replay cost kills savings. On-demand is increasingly the rational choice for production, despite 25-30% higher cost.

Should teams evaluate B200 now or wait for maturity?

B200 is mature enough for evaluation. Run a non-critical training job on B200 in the next 4 weeks. If results are solid, switch new projects to B200. If issues arise, stay on H100. The early feedback from production teams will inform pricing and availability through Q2.

Sources