Yes—but the original headline is now historical. Google announced at Cloud Next ’24 that NVIDIA’s Blackwell platform would arrive on Google Cloud in early 2025. The rollout happened in stages: A4 virtual machines with eight NVIDIA B200 GPUs entered preview on January 31, 2025, became generally available on March 18, and A4X machines based on NVIDIA GB200 NVL72 entered preview on February 19 before reaching general availability on May 29, 2025.
As of August 2026, Google Cloud lists both A4/B200 and A4X/GB200 machine series. They are not interchangeable: A4 is a conventional eight-GPU option for large-scale training and inference, while A4X targets models and workloads that can exploit rack-scale, tightly coupled infrastructure.
As an Amazon Associate I earn from qualifying purchases.
What Google actually announced
“Blackwell” referred to more than one Google Cloud product. Google’s April 9, 2024 announcement promised two configurations:
Free tools Windows power users keep installed
One-click scans. No signup required.
| Google Cloud offering | NVIDIA hardware | Typical use |
|---|---|---|
| A4 | Eight NVIDIA HGX B200 GPUs per VM | Foundation-model training, fine-tuning, inference, analytics and HPC |
| A4X | NVIDIA GB200 NVL72 rack-scale system | Very large models, reasoning, long-context inference and high concurrency |
That distinction matters because a B200 GPU and a GB200 system are not the same thing. GB200 combines Grace CPUs with Blackwell GPUs, and the NVL72 design connects 72 Blackwell GPUs through a high-bandwidth system-level interconnect.
#1 Best Overall
- Discrete graphics card memory 40 GB
- Memory bandwidth (max) 1555 GB/s
- Graphics processor family NVIDIA
- Graphics processor A100
Google’s original Cloud Next announcement described the expected arrival as “early 2025.” That wording meant a rollout window—not unrestricted general availability on January 1.
The rollout timeline
- April 9, 2024: Google announced plans to bring HGX B200 and GB200 NVL72 systems to Google Cloud in early 2025.
- January 31, 2025: A4 VMs powered by NVIDIA B200 GPUs entered preview.
- February 19, 2025: A4X VMs powered by NVIDIA GB200 entered preview.
- March 18, 2025: Google announced general availability for A4; A4X remained in preview at that point.
- May 29, 2025: Google’s A4X announcement page records the general-availability date for A4X.
Preview, general availability, and usable capacity are different milestones. Even after a machine type reaches general availability, customers may need quota approval, a reservation or another approved capacity arrangement, and access to a supported zone.
What an A4 VM provides
Google’s current accelerator-optimized machine documentation lists the main A4 configuration as a4-highgpu-8g. Its listed specifications include:
- Eight NVIDIA B200 GPUs
- 180 GB of GPU memory per GPU, or 1,440 GB in total
- 224 vCPUs
- 3,968 GB of host memory
- 12,000 GiB of local SSD
- Up to 3,600 Gbps of listed network bandwidth
- NVLink full-mesh GPU-to-GPU interconnect
That makes A4 the easier Blackwell option to compare with earlier multi-GPU cloud servers. It fits distributed training, fine-tuning, multi-GPU inference and other workloads that can use eight GPUs without requiring a complete rack-scale domain.
See Google’s accelerator-optimized machine documentation for current specifications and restrictions.
Rank #2
- Model: RTX 2000 ADA Generation
- Memory: 16GB GDDR6
- Satisfaction Ensured.
- Produced with the highest grade materials
- Memory: 16GB GDDR6
What an A4X VM provides
A4X is built around NVIDIA GB200 Grace Blackwell Superchips and the GB200 NVL72 architecture. Google lists the A4X machine type as a4x-highgpu-4g, with four GB200 superchips and four Blackwell GPUs per VM.
Current listed specifications include:
- 744 GB of total GPU memory per listed VM
- 140 vCPUs
- 884 GB of instance memory
- 12,000 GiB of local SSD
- Up to 2,000 Gbps of listed network bandwidth
The full NVL72 system links 72 Blackwell GPUs and 36 Grace CPUs through fifth-generation NVLink. Therefore, an A4X VM should not be described as a single 72-GPU virtual machine. The four-GPU VM is a unit within a much larger, tightly integrated system.
Recommended Free Tools
A4X is most compelling when software can exploit that scale-up topology—for example, very large foundation models, mixture-of-experts systems, reasoning models, long-context inference and high-concurrency serving. It is not automatically the better choice for every GPU workload.
Google’s A4X announcement explains the architecture and intended workloads.
A4 versus A4X
| Consideration | A4/B200 | A4X/GB200 |
|---|---|---|
| Basic unit | Eight-GPU VM | Four-GPU VM within an NVL72 system |
| Architecture | HGX B200 | GB200 Grace Blackwell Superchip and NVL72 |
| Best fit | Large training, fine-tuning, inference and HPC | Very large models and rack-scale workloads |
| Listed GPU memory | 180 GB per GPU; 1,440 GB per VM | 744 GB per listed VM |
| Listed maximum network bandwidth | 3,600 Gbps | 2,000 Gbps |
| Operational profile | More conventional multi-GPU deployment | More specialized scale-up deployment |
The key choice is scale-up versus scale-out. A4 resembles a large, conventional GPU VM. A4X is designed around communication and memory behavior across a much larger accelerator domain. A4X’s advantages may not appear if the model, parallelism strategy or framework cannot use that topology efficiently.
Rank #3
- 3328 optimized CUDA Cores, 7.99 TFLOPS
- 104 third generation Tensor Cores, 63.9 TFLOPS
- 26 third generation RT Cores, 15.6 TFLOPS
- Dual-slot width, low-profile form factor
- 70W maximum power consumption
What performance claims mean
Google reported that A4 delivered up to a 2.2-times training-performance improvement over A3 Mega in its cited comparison. It also reported 72 petaflops of FP8 performance for an eight-GPU A4 VM, 720 petaflops of FP8 performance for a complete NVL72 system, and 860,000 tokens per second for an inference test involving Llama 2 70B on a full NVL72 system.
These are vendor-reported figures, not universal or independent benchmarks. Results depend on model architecture, precision, batch size, software versions, parallelism, input length and whether the comparison concerns one VM or a complete NVL72 system. They should not be converted into a claim that A4 is 2.2 times faster for every AI workload.
Google presents Blackwell as part of its broader AI Hypercomputer architecture, which also includes networking, storage, scheduling, GKE, Slurm, Cluster Director and Titanium infrastructure offloads. In practice, data loading, checkpointing, placement, collective communication and failure recovery can matter as much as peak GPU specifications.
Availability and provisioning constraints
A4 and A4X are not equivalent to ordinary, instantly available VM types. Availability varies by region and zone, and a customer may need quota approval, a reservation or another capacity mechanism. Google’s live GPU regions and zones table should be checked before deployment because locations change.
Google’s documentation also lists important restrictions:
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteRank #4
- NVIDIA Ampere Streaming Multiprocessors: Building blocks for the world's fastest, most efficient GPUs, the all-new Ampere SM brings twice the FP32 throughput and improved energy efficiency
- 2nd Generation RT Cores - Experience 2x the 1st Generation RT Cores throughput, plus competitive RT and shading for a whole new level of ray-tracing performance
- 【3rd Generation Tensor Cores】Get up to 2X the throughput with structural sparsity and advanced AI algorithms such as DLSS
- Core Clock: 1837MHz
- WINDFORCE 3X Cooler
- A4 and A4X are available only in selected zones.
- Hyperdisk is required; Persistent Disk cannot be used.
- Neither machine series supports Windows.
- A4X does not support Spot VMs, Flex-start VMs or sole tenancy.
- A4X does not receive sustained-use discounts or flexible committed-use discounts.
- The Compute Engine SLA does not apply to the A4X machine series.
These limitations can require changes to infrastructure-as-code templates, storage design, operating-system assumptions and scheduling strategy. Review the GPU VM creation and limitation documentation before treating a published machine type as immediately usable.
Pricing is more complicated than an hourly GPU number
Google’s accelerator-optimized pricing page showed the following A4 eight-B200 pricing signals when checked in August 2026:
| Consumption mode | Displayed hourly figure |
|---|---|
| On-demand | N/A |
| Dynamic Workload Scheduler Flex-start | $64.4400 |
| Dynamic Workload Scheduler Calendar Mode | $90.22 |
| Current Spot | $37.7464 |
| One-year compute resource commitment | $88.9272 |
| Three-year compute resource commitment | $56.7072 |
These are dated pricing signals, not guaranteed universal rates. Google pricing varies by region and consumption model, and the displayed figures do not make A4 equivalent to a normal on-demand VM rental. A4X’s consumption restrictions also make it unsuitable for buyers seeking ordinary elastic or preemptible economics.
The complete bill can include VM resources, local SSD, Hyperdisk, networking, storage, data transfer and orchestration or cluster-management costs. Check the accelerator-optimized pricing page, GPU pricing page and Google Cloud pricing calculator immediately before committing capacity.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →When Blackwell is the wrong choice
The newest accelerator is not automatically the cheapest or simplest option. A4 or A4X may be a poor fit when:
Best Value
- 900-1G136-2505-000
- The model is small enough for L4, T4, A100, H100 or H200 capacity.
- The workload is bursty and cannot use reserved capacity efficiently.
- The application cannot keep eight GPUs busy.
- The software stack has not been validated against the required CUDA, driver, framework and distributed-training versions.
- The workload requires Windows or Persistent Disk.
- The organization needs Spot or Flex-start economics on A4X.
- Capacity is unavailable in the required region or data-residency boundary.
- The workload would be cheaper or simpler on a Google TPU.
Alternatives worth comparing
A3 Ultra and A3 Mega
Google Cloud’s A3 Ultra H200 and A3 Mega H100 machines remain relevant alternatives. They may be preferable when a team already has mature H100 or H200 images, benchmarks and deployment tooling, or when their capacity and pricing are more favorable. A4 should not be selected solely because it is newer.
Google Cloud TPUs
TPUs can be attractive for workloads already built around JAX, TensorFlow or Google’s TPU-optimized software stack. GPUs generally provide broader portability across CUDA-based tools, while TPUs may offer better economics for suitable models. The right choice depends on software fit and measured cost per useful result.
Google’s TPU documentation is the appropriate starting point for that comparison.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
G4 and RTX PRO 6000 Blackwell
G4 VMs use NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and target visual computing, graphics, fine-tuning and inference. They are not a substitute for the B200/GB200 architecture when the requirement is rack-scale distributed training.
See Google’s G4 announcement for its intended positioning.
What to benchmark before buying
A meaningful pilot should measure the workload rather than rely on peak FP8 figures. Test:
- Tokens per second at the production batch size
- Time to first token and inter-token latency
- Training throughput per dollar
- Scaling efficiency from one VM to multiple VMs
- Checkpoint and restart time
- Data-loading throughput and storage cost
- GPU memory headroom and actual utilization
- NCCL or equivalent collective-communication performance
- Fault recovery and interruption behavior
- Quota approval and capacity lead time
- Regional availability and data-transfer costs
- CUDA, PyTorch, JAX, Kubernetes and Slurm compatibility
A peak-throughput test can be misleading for memory-bound, communication-bound, retrieval-heavy or irregular workloads. The useful metric is usually completed training, generated tokens or served requests per total dollar—not the accelerator’s theoretical maximum.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchQuick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
