Google Cloud A4 B200 Pricing 2026: GCP Cost vs Spheron

TL;DR: GCP A4 B200 vs Spheron B200 (6 Jul 2026)

Pricing fluctuates based on GPU availability. The prices above are based on 6 Jul 2026 and may have changed. Check current GPU pricing → for live rates. GCP figures are cross-referenced third-party pricing trackers as of the same date; verify with the GCP pricing calculator before budgeting.

A4 is Google's newest GPU machine family, built around NVIDIA's B200, and it's a genuinely different instance line from the A3 High series this blog already covers for H100. It also undercuts AWS's B200 instance by a wide margin. Where it gets more interesting, and less flattering to us, is the comparison against Spheron's own live B200 rate: right now Spheron only lists B200 as spot capacity, and at full 8-GPU node scale GCP's guaranteed on-demand A4 rate actually comes in lower. We'll walk through exactly where each option wins.

For the broader Blackwell pricing picture across providers, see our B200 cloud pricing comparison and the GPU cloud pricing hub covering 5+ providers.

What A4 VMs Are and How They Differ From A3 H100 Instances

A4 VMs pack 8x NVIDIA B200 (Blackwell) GPUs connected over 5th-generation NVLink, and they're a distinct machine family from A3 High's 8x H100 SXM5 setup, not a spec bump on the same line. Google made that separation explicit when it introduced A4, and the numbers back it up: each B200 in an A4 VM delivers 2.25x the peak compute and 2.25x the HBM capacity of the H100 in A3 High, according to Google's own announcement.

"The Blackwell architecture represents a giant step forward for the AI industry, so we're excited that the B200 GPU is now available with the new A4 VM," said Ian Buck, VP and General Manager of Hyperscale and HPC at NVIDIA, at the launch.

A4 vs A3 High Spec Comparison

Specs from Google Cloud's GPU machine type docs. The per-GPU memory jump (80 GB to 180 GB) matters more in practice than the headline compute multiplier: it's the difference between needing two H100s to hold a 70B model in FP16 and fitting it comfortably on one B200.

A4's networking is also a real step up, not just a bigger number. The VM connects GPUs at a non-blocking 3.2 Tbps over RoCE through Google's Titanium ML network adapter (a ConnectX-7-based NIC), on top of Google's Jupiter datacenter fabric, rated at 13 petabits/sec of bi-sectional bandwidth for scaling to tens of thousands of GPUs, per Google's announcement. For multi-node training jobs where GPU-to-GPU communication is the bottleneck, that bandwidth is the actual product being sold, not just a spec sheet line.

Why Google Moved to a Separate Instance Family for Blackwell

Google didn't ship A4 as "A3 Ultra" or a bigger A3 SKU. It's its own machine series with its own naming, its own regional rollout, and its own reservation model. Part of that is architectural: Blackwell's NVLink generation, memory type, and network adapter are different enough from Hopper that reusing the A3 chassis design wasn't practical. Part of it is commercial: keeping A4 separate lets Google price and allocate B200 capacity independently of the (now heavily discounted, widely available) A3 H100 fleet, and route it through the newer AI Hypercomputer reservation system rather than retrofitting the older CUD machinery A3 already uses.

A4 VMs reached general availability on March 17, 2025, per Google Cloud's Compute Engine release notes, a little over 16 months before this post. As of this writing, A4 is available in 11 zones across the US, Europe, and Asia, per Google's GPU regions and zones documentation: Tokyo, Singapore, Hamina, Eemshaven, Council Bluffs, Moncks Corner, Ashburn, Dallas, Los Angeles, and Salt Lake City. Despite that head start, A4's footprint is still narrower than A3 High's, which has had years to roll out across GCP's full region list.

A4 On-Demand Pricing Breakdown

The a4-highgpu-8g list price runs approximately $34.24/hr on-demand in us-central1, per independent GCP pricing trackers DevZero and CloudPrice, cross-checked against each other since Google doesn't render a static price table for A4 on its public pricing pages. Treat that number as directional and confirm it against the GCP pricing calculator before signing off on a budget.

Per-GPU Math on the a4-highgpu-8g List Price

Divide the node price by 8 GPUs:

$34.24/hr ÷ 8 = $4.28/hr per B200 GPU

That's the number to compare against other providers' per-GPU B200 rates, not the $34.24 headline, since every competing offer in this post is normalized to a per-GPU basis too.

Regional Price Variance

Google typically prices GPU capacity within roughly 5-15% of us-central1 across its other US and EU regions, based on the pattern already documented for A3 High. Independent trackers don't yet publish a full regional price table specifically for A4 broken out by all 11 zones, so we're not going to invent region-by-region deltas here. If your workload is tied to a specific zone (europe-west4-b for EU data residency, for example), run that zone through the pricing calculator directly rather than assuming a flat percentage markup.

Committed-Use, Reservations, and Spot for A4

Why Standard 1yr/3yr CUDs Don't Apply to A4

If you've priced A3 H100 or older GCP GPU instances before, you're used to the standard playbook: commit for 1 or 3 years, get 20-46% off. That playbook doesn't exist for A4. Google's own committed-use discount documentation states that CUD recommendations are not available for the A3 Ultra, A4, A4X, and A4X Max machine series. Instead, discounted, guaranteed A4 capacity runs through a separate mechanism: future reservations under the AI Hypercomputer consumption option, which requires contacting your Google account team or the sales team before you can submit a request. There's no self-service equivalent of clicking "buy a 3-year CUD" in the console for A4 the way there is for A3.

Practically, this means the $4.28/hr per-GPU figure above is close to the real floor for self-service, on-demand A4 access unless you're negotiating a capacity deal directly with Google. Budget accordingly if your model assumed a CUD discount would bring that number down further on its own.

A4 Spot VM Availability

Spot VMs are supported for A4, unlike A4X (which Google's Spot VM documentation explicitly excludes). Google has published Spot/Preemptible SKU groups for A4 across three dozen-plus regions worldwide, which confirms the billing plumbing exists even though per-hour spot rates aren't broken out on a simple public price list. Spot discounts on GCP typically run 60-91% off on-demand depending on machine type and region, but expect that range to run narrower for a GPU as scarce and in-demand as the B200. As with any spot capacity, GCP can reclaim an A4 Spot VM at any time, so it's a fit for checkpointed training and batch jobs, not production inference serving.

Hidden Costs on GCP A4

The $34.24/hr headline doesn't cover everything you'll actually pay. The same cost categories that inflate GCP A3 bills apply to A4, often at a larger absolute scale because the node itself is more expensive:

• Data egress. $0.08-$0.12/GB for data leaving the GCP VPC. A4's larger per-GPU memory means larger checkpoints by default (a 70B model checkpoint that needed two H100s to hold now fits on one B200, but the file on disk is the same size or bigger once you add optimizer states), so a training run that regularly ships checkpoints out of GCP will rack up more egress in absolute dollars than the same job on A3.
• Persistent disk. A4's local SSD is ephemeral, same as A3. Standard PD runs about $0.04/GB/month, SSD PD about $0.17/GB/month. Ten terabytes of checkpoint storage on SSD PD is roughly $1,700/month on top of compute.
• Quota and reservation friction. New GCP accounts default to zero vCPU quota for GPU-heavy machine families, and A4's newer, scarcer capacity makes this worse in practice than A3. Getting an A4 quota increase or an AI Hypercomputer future reservation approved can take a written business justification and a multi-day (sometimes multi-week) wait, which is a real cost if your project timeline assumes same-day access.
• The node-lock-in tax. This one is specific to A4. Because a4-highgpu-8g is the only shape Google sells, a project that needs a single B200 for experimentation still has to provision, and pay for, all 8 GPUs. There's no a4-highgpu-1g the way there's an a3-highgpu-1g. If your actual workload needs one or two GPUs, the effective per-GPU cost of the A4 path is whatever fraction of $34.24/hr you actually use divided by however many GPUs you needed, which can be far worse than $4.28/hr in practice.

For a broader look at hidden hyperscaler costs across AWS, GCP, and Azure, see our hyperscaler cost breakdown.

A4 vs Spheron B200: Cost Per GPU-Hour at Scale

Here's where we owe you the honest version rather than the flattering one. Pulled live from Spheron's pricing API on 6 July 2026, B200 SXM6 is currently listed only as spot capacity: $5.37/hr for a single GPU, and $42.70/hr for an 8-GPU node ($5.34/hr per GPU). There is no dedicated, on-demand B200 tier live on Spheron right now. That's a real, checkable state of the market, not a rounding artifact: we split the API response by instanceType before computing anything, per our own pricing methodology, and every current B200 offer comes back SPOT.

That changes the comparison. On a pure per-GPU-hour basis at full node scale, GCP A4's guaranteed on-demand rate of $4.28/hr per GPU is actually about 20% cheaper than Spheron's current $5.34/hr spot floor for an 8-GPU B200 node, and it comes with none of the preemption risk that spot capacity carries. If your workload needs guaranteed, non-interruptible B200 access at 8-GPU scale today, GCP A4 is the better price right now, not Spheron.

Single-GPU and 8-GPU Node Cost Comparison

GCP A4's on-demand rate beats AWS's B200 instance by roughly 3.3x per GPU-hour ($4.28 vs $14.24), which is the headline number this post set out to check, and that comparison holds up. It's the Spheron column where the story is more nuanced this month: we're cheaper for anyone who needs a single GPU (GCP simply won't sell you one), and once B200 comes back online as an on-demand tier on Spheron the full-node math will likely flip back in our favor the way it does for H100 and H100-class comparisons. Right now, for guaranteed 8-GPU B200 capacity, GCP A4 is the cheaper of the two.

If your workload can tolerate interruption, checkpointed multi-day training being the classic case, renting B200 GPUs on Spheron still lets you buy exactly the GPU count you need instead of a fixed 8-GPU block, which is worth real money if your actual requirement is one or two GPUs. Check current B200 pricing for the live spot rate before committing.

Where GCP A4 Still Makes Sense

• You need guaranteed 8-GPU B200 capacity today. At $4.28/hr per GPU, A4 on-demand undercuts both AWS's B200 instance and Spheron's current spot rate, with no preemption risk.
• Deep Vertex AI or GKE integration. If your training and serving pipeline already runs on Vertex AI Pipelines or GKE with GPU node pools, A4 slots directly in without re-architecting deployment tooling.
• BigQuery-adjacent data pipelines. Training loops reading large datasets straight out of BigQuery avoid cross-cloud egress entirely by staying on GCP.
• You can secure an AI Hypercomputer future reservation. Teams with an existing Google account relationship and a large enough footprint to justify a reservation conversation can likely negotiate below the $4.28/hr list price.

Where GCP A4 Falls Short

• You need fewer than 8 GPUs. There's no smaller A4 shape. A single-GPU experiment on GCP means renting, and paying for, the whole node.
• You want spot pricing without a sales conversation. A4 Spot VMs exist as a billing mechanism, but the exact rate isn't published on a simple price list the way Spheron's is.
• You want to avoid quota and reservation friction. New accounts start at zero GPU quota, and A4's constrained regional footprint makes approval slower than for A3.

Monthly Cost Scenarios

Scenario 1: Solo Researcher, 1x B200, 160 hrs/month

A researcher running weekly experiments on a single B200 for a month:

• Spheron (1x B200 spot, 160 hrs): 160 × $5.373 = $859.68
• GCP A4: no single-GPU shape exists. Getting any B200 access at all means renting the full a4-highgpu-8g node: 160 × $34.24 = $5,478.40, with 7 of 8 GPUs sitting idle the entire time.

For anyone who genuinely needs one GPU, GCP A4's node-lock-in makes it roughly 6.4x more expensive than Spheron in this scenario, not because the underlying rate is bad, but because you can't buy less than 8.

Scenario 2: Small Team, 24/7 Inference Needing 2 GPUs of Capacity

A team running production inference that occupies roughly 2 B200s worth of throughput, continuously, for a month (720 hours):

• GCP A4 (full 8-GPU node, on-demand, guaranteed): 720 × $34.24 = $24,652.80, using only 2 of the 8 provisioned GPUs but with zero preemption risk for a production workload.
• Spheron: no on-demand B200 tier is currently live, so matching this workload today means either accepting spot preemption risk for production traffic (not recommended) or provisioning 2 GPUs' worth of spot capacity at 720 × 2 × $5.373 = $7,737.12 and building interruption-handling into the serving layer.

This is the scenario where the honest tradeoff shows up clearest: Spheron is far cheaper in raw dollars, but GCP A4 is the safer default for production traffic that can't tolerate a mid-request preemption, until Spheron's B200 on-demand tier returns.

Scenario 3: 8-GPU Training Cluster, 2-Week Checkpointed Run (336 hrs)

A team running a 2-week distributed training job on a full 8-GPU B200 cluster, with checkpointing so interruptions are recoverable:

• GCP A4 (on-demand, full node): 336 × $34.24 = $11,504.64
• Spheron (8-GPU node, spot): 336 × $42.7005 = $14,347.37
• AWS p6.48xlarge (on-demand): 336 × $113.93 = $38,280.48

For this fully node-matched, checkpoint-tolerant scenario, GCP A4 currently comes in about 20% cheaper than Spheron's spot rate and about 70% cheaper than AWS on-demand. That's the clearest illustration of the point this post is built around: A4 is a genuinely well-priced instance relative to AWS, and right now it also beats our own spot rate at full 8-GPU scale. We'd rather tell you that directly than pretend otherwise.

Pricing fluctuates based on GPU availability. The prices above are based on 6 Jul 2026 and may have changed. Check current GPU pricing → for live rates. GCP figures come from independent pricing trackers cross-referenced against each other; verify with the GCP pricing calculator before budgeting. For setup and deployment details on Spheron, see docs.spheron.ai.

If your B200 workload fits in a single GPU or needs the flexibility to scale by one GPU at a time instead of committing to a fixed 8-GPU node, that's exactly the gap Spheron's marketplace is built to fill.

Check B200 pricing on Spheron →