AWS and Google Cloud are no longer fighting for the same growth curve. Amazon still runs the biggest cloud on the planet by a wide margin, but Google Cloud posted 82% year-over-year revenue growth in its most recent quarter against AWS’s 37%, according to earnings breakdowns reported by The Motley Fool on September 9, 2026. That gap is reshaping how enterprise buyers weigh the two platforms heading into 2027 budget cycles. This comparison breaks down the September 2026 numbers that actually matter: market share, the new AWS Graviton5 and Google Axion N4A chips, storage and egress pricing, Bedrock versus Vertex AI token costs, and Kubernetes control-plane fees, plus a migration path if you’re moving workloads between the two.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
AWS vs Google Cloud in September 2026: Where the Market Stands
Synergy Research Group’s Q2 2026 tracking, cited across multiple industry summaries, puts AWS at 28% of global cloud infrastructure spending, down from roughly 30% a year earlier. Google Cloud climbed to 15%, its highest share on record and up from about 13% in the same period last year. Microsoft Azure sits at 20% in the same dataset, meaning the three hyperscalers combined still control 63% of the market. The share numbers only tell part of the story, though. AWS still generates more absolute revenue than Google Cloud by a wide margin: AWS closed the quarter at $42.2 billion in revenue with 37% year-over-year growth, while Google Cloud’s growth rate hit 82%, according to the Fool’s earnings recap.
That growth differential, first flagged when Google Cloud’s growth hit 82% ahead of AWS’s own earnings report, matters for buyers because it signals where the two companies are investing. AWS is defending an enormous existing footprint, spreading spend across more than 30 regions and a catalog of services that dwarfs any competitor. Google Cloud is growing off a smaller base but is converting that growth into faster iteration on newer categories, particularly AI infrastructure, ARM compute, and BigQuery-adjacent analytics tooling. Neither trend is new, but the size of the growth gap in 2026 is the widest it has been since Google Cloud became a reportable Alphabet segment. For teams comparing AWS vs Google Cloud today, the practical takeaway is that AWS remains the safer default for breadth and maturity, while Google Cloud is where a disproportionate share of new-generation compute and AI tooling is shipping first.
It’s worth being precise about what that year-over-year shift actually represents. AWS’s move from roughly 30% to 28% share and Google Cloud’s move from roughly 13% to 15% share, both measured over the same trailing twelve months by Synergy Research, happened while AWS’s absolute revenue kept growing at 37% year-over-year. That’s not a shrinking business, it’s a business growing more slowly than a rapidly expanding total market. Google Cloud’s 82% growth rate is capturing a disproportionate share of new cloud spend rather than pulling existing AWS workloads away wholesale, which is the more common (and slower) way market share actually shifts between entrenched hyperscalers. For a buyer, the practical implication is that AWS isn’t losing ground on reliability or feature depth, it’s losing ground on the rate at which new AI and analytics workloads are choosing Google Cloud as their starting point.
| Category | AWS | Google Cloud |
|---|---|---|
| Global cloud infrastructure share (Q2 2026) | 28% | 15% |
| Year-over-year revenue growth | 37% | 82% |
| Latest quarterly cloud revenue | $42.2 billion | Not separately disclosed in the same breakdown |
| Latest custom ARM chip | Graviton5 (R9g/R9gd, M9g/M9gd) | Axion N4A (newest) / C4A |
| Chip process node | 3nm (TSMC) | Not publicly disclosed for N4A |
| Chip cores (top instance) | 192 cores per chip | Up to 72 vCPUs (C4A), 64 vCPUs (N4A) |
| Flagship LLM on the platform | Amazon Nova Premier / Claude Opus 4.1 via Bedrock | Gemini 3.1 Pro via Vertex AI |
| Flagship LLM output price (per 1M tokens) | $12.50 (Nova Premier) / $75 (Claude Opus 4.1) | $12 (Gemini 3.1 Pro, up to 200K context) |
| Managed Kubernetes control plane | $0.10/cluster-hour (EKS Standard) | $0.10/cluster-hour (GKE Standard) |
| Object storage starting price | $0.023/GB-month (S3 Standard, first 50TB) | $0.020/GB-month (Cloud Storage Standard, single region) |
| Internet egress starting price | $0.09/GB (after 100GB free) | $0.085/GB Standard Tier / $0.12/GB Premium Tier (after ~200GB free) |
| Customers on custom ARM silicon | 120,000+ (AWS Graviton family) | Not separately disclosed for Axion |
AWS Graviton5 vs Google Axion: The New ARM Chip Race
The most consequential product news in the AWS vs Google Cloud fight this quarter isn’t a new region or a price cut, it’s silicon. AWS pushed Amazon EC2 R9g and R9gd instances, powered by AWS Graviton5, to general availability in late August 2026, according to the AWS News Blog. Graviton5 packs 192 cores across a four-chiplet design built on Arm’s Neoverse V3 core and the Armv9.2-A instruction set, fabricated on a 3-nanometer TSMC process. AWS says the chip delivers up to 25% better overall compute performance than the previous-generation Graviton4 (R8g), with up to 35% faster web application performance, 35% faster machine learning inference, and 30% faster database throughput, building on the price and performance gap AWS’s own Graviton line has already opened up against Intel and AMD x86 instances. It also ships with DDR5-8800 memory, which AWS describes as the fastest DDR5 currently deployed in any cloud instance, along with a 192MB L3 cache, five times larger than Graviton4’s.
Google’s answer is the Axion family, and it now spans two generations available at once. C4A, the original Axion chip built on Arm Neoverse V2 cores, has been generally available since 2025 and scales up to 72 vCPUs with 576GB of DDR5 memory and 100 Gbps of networking. N4A, the newer general-purpose Axion line, reached general availability in 2026 and targets cost-sensitive, scale-out workloads with up to 64 vCPUs and 512GB of memory. Google’s own numbers, published on the Google Cloud blog, claim N4A delivers up to 105% better price-performance than comparable current-generation x86 VMs for compute-bound workloads, 90% for scale-out web servers, and 85% for Java applications. C4A’s own claim is more modest but still notable: up to 65% better price-performance and 60% better energy efficiency versus comparable x86 instances, per the Google Axion product page.
Both companies are making Arm-vs-x86 claims rather than Arm-vs-Arm claims, which makes a direct Graviton5-vs-Axion comparison harder to pin down from vendor material alone. That’s where independent benchmarking fills the gap, and the results are covered in the next section. What’s clear without any third-party data is that AWS currently ships the higher raw core count (192 vs Axion’s 72), while Google’s N4A carries the more aggressive price-performance claim of the two 2026-era chips.
| Spec | AWS Graviton5 (R9g/R9gd) | Google Axion C4A | Google Axion N4A |
|---|---|---|---|
| Core architecture | Arm Neoverse V3, Armv9.2-A | Arm Neoverse V2 | Arm Neoverse V2 (refined) |
| Process node | 3nm (TSMC) | Not publicly disclosed | Not publicly disclosed |
| Max cores/vCPUs per chip | 192 cores (4 chiplets) | Up to 72 vCPUs | Up to 64 vCPUs |
| Max memory | DDR5-8800, up to 768GB (r9gd.24xlarge) | 576GB DDR5 | 512GB DDR5 |
| L3 cache | 192MB (5x Graviton4) | Not disclosed | Not disclosed |
| Max network bandwidth | Up to 17 Gbps (r9g) | 100 Gbps | 50 Gbps |
| Clock speed | 3.3GHz (per Phoronix testing) | Not disclosed | Not disclosed |
| GA date | August 2026 | 2025 | 2026 |
| Vendor price-performance claim vs x86 | Not framed as x86 comparison (25-35% vs Graviton4) | Up to 65% better, 60% more energy efficient | Up to 105% better (compute-bound) |
| PCIe generation | Gen 6 | Not disclosed | Not disclosed |
Compute Pricing: R9g vs N4A vs C4A Instance Costs
List pricing for the two chip families is not a clean apples-to-apples match because AWS and Google Cloud size their instance families differently. AWS’s r9g.4xlarge, a 16 vCPU, 128GB memory instance running Graviton5, lists at $1.0274 per hour on-demand in us-east-1. Google’s smallest published N4A configuration, n4a-standard-1, runs a single vCPU at $0.0385 per hour in us-central1. Scaling that per-vCPU rate linearly to a comparable 16-vCPU N4A instance would land in the neighborhood of $0.60 to $0.65 per hour, though Google has not published that exact configuration’s price directly, so treat that figure as an estimate rather than a confirmed list price.
The pricing gap matters less in isolation and more in context of the price-performance claims each vendor makes. If Google’s N4A really does deliver up to 105% better price-performance than comparable x86 instances for compute-bound work, and AWS Graviton5 delivers “only” 25% better compute than Graviton4 (its own prior generation, not an x86 baseline), then a workload that’s genuinely scale-out and stateless may see a larger effective cost reduction on Google Cloud’s newest Arm instances than on AWS’s newest Arm instances, purely because of how each vendor is framing and pricing the improvement. That’s a reason to run your own workload-specific benchmark rather than trust either vendor’s blog post as a substitute for a proof of concept.
| Service | AWS Price | Google Cloud Price |
|---|---|---|
| ARM compute, entry instance | r9g.4xlarge (16 vCPU/128GB): $1.0274/hr | n4a-standard-1 (1 vCPU): $0.0385/hr |
| Object storage, Standard tier | $0.023/GB-month (first 50TB, us-east-1) | $0.020/GB-month (single region) |
| Internet egress, entry tier | $0.09/GB (0-10TB, after 100GB free) | $0.085/GB Standard Tier (0-10TiB, after ~200GB free) / $0.12/GB Premium Tier (0-1TB) |
| Flagship native LLM, output tokens | $12.50/1M (Amazon Nova Premier) | $12/1M, ≤200K context (Gemini 3.1 Pro) |
| Top-tier frontier model, output tokens | $75/1M (Claude Opus 4.1 via Bedrock) | $18/1M, >200K context (Gemini 3.1 Pro) |
| Managed Kubernetes control plane | $0.10/cluster-hour (EKS Standard, ~$73/mo) | $0.10/cluster-hour (GKE Standard, ~$73/mo, 1 free cluster) |
Benchmark Results: What Independent Testing Shows
Vendor marketing claims are a starting point, not a verdict. Three independent benchmarking efforts published in 2026 give a clearer picture of how Graviton5 and Axion actually perform. Phoronix’s Graviton5 review measured a 30% geometric-mean improvement over Graviton4 across its standard Linux benchmark suite, with a measured clock speed of 3.3GHz, broadly consistent with AWS’s own 25% headline figure. Independent testing from SpareCores measured Graviton5 against Graviton4 directly and found a 16.5% single-core performance gain, a 17.5% multi-core gain (40% on PassMark’s CPU Mark specifically), and a memory latency improvement from 48.88ms down to 30.71ms.
On the database side, GizmoData ran a TPC-H 1TB benchmark, a standard analytical query workload, across matched instance sizes. The result: an r9gd.24xlarge (Graviton5) completed the 22-query benchmark in 80.475 seconds, versus 89.831 seconds for the equivalent r8gd.24xlarge (Graviton4), a real-world confirmation of AWS’s database performance claims that lines up closely with the vendor’s own “up to 30% faster for databases” figure once cache effects are accounted for.
For Google’s side, DoiT’s independent benchmarking, published as “First Look at Google Cloud N4A VMs,” tested N4A against both its C4A predecessor and AWS’s M8g Graviton-based instance. N4A came out ahead in raw CPU throughput on both counts, outperforming C4A by 13.7% and AWS’s M8g by 21.7%, according to DoiT’s published numbers. That’s a genuinely useful data point because it’s one of the few tests that puts a Google chip and an AWS chip side by side rather than comparing each against its own prior generation or against generic x86. The caveat is that M8g is a Graviton4-generation instance, not Graviton5, so a fresh head-to-head against R9g/M9g hadn’t been published as of this writing.
| Benchmark source | What was tested | Result |
|---|---|---|
| Phoronix | Graviton5 vs Graviton4, full Linux benchmark suite | 30% geomean improvement, 3.3GHz clock confirmed |
| SpareCores | Graviton5 vs Graviton4, single/multi-core and memory latency | +16.5% single-core, +17.5%/+40% multi-core (PassMark), latency cut from 48.88ms to 30.71ms |
| GizmoData | TPC-H 1TB, r9gd.24xlarge vs r8gd.24xlarge vs Azure E96pds_v6 | 80.475s vs 89.831s vs 106.806s |
| DoiT | Google Axion N4A vs C4A vs AWS M8g, raw CPU throughput | N4A beats C4A by 13.7%, beats AWS M8g by 21.7% |
Storage Pricing: S3 vs Google Cloud Storage
Object storage pricing is one of the few areas where the AWS vs Google Cloud gap is small and stable rather than volatile. AWS S3 Standard in us-east-1 lists at $0.023 per GB per month for the first 50TB, stepping down to $0.022/GB for the next 450TB and $0.021/GB beyond 500TB. Google Cloud Storage Standard, single-region pricing in a comparable US region, lists at $0.020 per GB per month, a roughly 13% discount against AWS’s entry tier before any volume discounts kick in on either side.
Neither number is the whole story for a real bill. Both platforms charge separately for requests (PUT/GET operations), for retrieval on cooler storage classes like S3 Glacier or Google Cloud Storage Coldline/Archive, and for cross-region replication. At small scale, under a few terabytes, the difference between $0.023 and $0.020 per GB is close to rounding error. At the scale where it starts to matter, tens or hundreds of terabytes, the discount compounds against a much larger base, and Google Cloud’s flatter tiering (no volume breakpoints published as of this writing) versus AWS’s stepped discount structure means the actual crossover point depends heavily on how much data you’re storing and how AWS’s own volume tiers apply to your account.
Data Egress Costs: The Hidden Bill
Egress is where cloud bills quietly balloon, and it’s also where AWS and Google Cloud diverge the most in structure. AWS gives every account 100GB of free internet egress per month, then charges $0.09/GB for the next 10TB, stepping down to $0.085/GB (10-50TB), $0.07/GB (50-150TB), and $0.05/GB beyond 150TB. Google Cloud splits its egress into two tiers. Premium Tier, which routes traffic over Google’s private global network for lower latency, starts at $0.12/GB for the first 1TB to North America, dropping to $0.11/GB (1-10TB) and $0.08/GB above 10TB. Standard Tier, which is cheaper but more latency-tolerant, gives roughly 200GB free per month, then charges $0.085/GB up to 10TiB, $0.065/GB up to 150TiB, and $0.045/GB beyond that.
At entry-tier volume, AWS’s $0.09/GB undercuts Google’s Premium Tier $0.12/GB by 25%, but Google’s Standard Tier at $0.085/GB is marginally cheaper than AWS at the same volume band. The practical decision for teams with heavy egress, media delivery, API backends serving large payloads, or multi-region replication, is whether the latency benefit of Google’s Premium Tier routing is worth the 33% premium over its own Standard Tier, and whether AWS’s simpler single-tier structure is easier to forecast even if it isn’t always the cheapest option at every volume level.
AI and Machine Learning: Bedrock vs Vertex AI
Both platforms have restructured their managed AI services around a multi-model marketplace rather than a single house model, but the flagship pricing tells a slightly different story on each side. Amazon Bedrock lists Amazon’s own Nova Premier as its native flagship: $2.50 per million input tokens and $12.50 per million output tokens, with a 1-million-token context window. Bedrock also hosts third-party frontier models, including Claude Opus 4.1 at $75 per million output tokens, and Claude Sonnet 5, which launched on Bedrock at promotional pricing of $2/$10 per million input/output tokens through August 31, 2026, before stepping up to standard $3/$15 pricing in September. xAI’s Grok 4.6 is also available on Bedrock at $2 per million input tokens and $6 per million output tokens.
Google Vertex AI centers its flagship pricing on Gemini 3.1 Pro: $2 per million input tokens and $12 per million output tokens for context windows up to 200,000 tokens, rising to $4 input / $18 output per million tokens above that threshold. Google is also running introductory pricing on its faster, cheaper Gemini 3.8/3.7/3.6 Flash models at $0.75/$3.75 per million input/output tokens through the end of December 2026, after which that pricing is expected to roughly double. Vertex AI also hosts Claude models directly, with Claude Fable 5.1 priced at $50 per million output tokens on that platform, distinct from AWS Bedrock’s own pricing for comparable third-party models.
The headline comparison, native flagship to native flagship, is close: Nova Premier’s $12.50 output price against Gemini 3.1 Pro’s $12 output price (at the smaller context tier) is nearly a wash. Where the two diverge is at the high end. AWS routes enterprise customers who want frontier-model quality toward Claude Opus 4.1 at $75/million output tokens through Bedrock, a premium tier Google doesn’t have an exact equivalent for inside Vertex AI’s own native lineup. For teams optimizing purely on cost per token at moderate context lengths, Google’s Gemini 3.1 Pro and its aggressively priced Flash tier currently undercut AWS’s comparable options, but the gap narrows once you factor in AWS’s broader third-party model marketplace inside Bedrock.
Kubernetes: EKS vs GKE
Managed Kubernetes pricing has converged almost completely between the two platforms at the baseline. Amazon EKS charges $0.10 per cluster-hour for Standard support, roughly $73 per month per cluster, with an Extended Support tier at $0.60 per cluster-hour for clusters running versions past their standard support window, and newer scaled control-plane tiers (XL through 8XL) priced from $1.65 up to $13.90 per cluster-hour for very large clusters that need dedicated control-plane capacity. Google Kubernetes Engine charges the identical $0.10 per cluster-hour for Standard mode, also about $73 per month, but includes one free zonal or Autopilot cluster per billing account, which effectively zeroes out the control-plane fee for smaller organizations running a single production cluster.
The control-plane fee was never where the real Kubernetes cost difference lived anyway, it’s in worker node pricing, autoscaling behavior, and how each platform handles spot/preemptible capacity. Google originated Kubernetes internally and GKE still tends to ship new upstream Kubernetes versions and features (Gateway API, multi-cluster services) slightly ahead of EKS in practice, though AWS has closed most of that gap over the past two years. For teams already deep in the AWS ecosystem, running EKS alongside existing IAM, VPC, and Bedrock integrations is usually simpler than bolting GKE onto an AWS-centric account. For greenfield cloud-native builds with no existing lock-in, GKE’s free cluster and Autopilot’s fully managed node provisioning remove a layer of operational overhead that EKS still requires you to manage yourself, or via a third-party add-on like Karpenter.
Real-World Deployments: Who’s Running What
Vendor benchmarks are useful, but production deployments are the real proof. On the Google Cloud side, Snowflake has adopted Axion-based infrastructure for its Gen2 warehouses, and Snowflake’s own published data credits Axion’s DDR5 memory architecture with up to a 50% improvement in memory bandwidth, which directly cuts query latency for hash-table and Bloom-filter-heavy operations common in data warehousing. ClickHouse Cloud reports 30-55% faster queries and roughly 15% fewer compute credits consumed after moving workloads onto Axion C4A instances, a concrete efficiency gain for a company whose entire product is query performance. Google’s own Cloud SQL team reports that Cloud SQL Enterprise Plus running on Axion-based C4A delivers up to 48% better price-performance than the previous-generation N2 machine family for transactional workloads, while Cloud SQL Enterprise on N4 machines shows a 44% improvement for the same workload class.
On the AWS side, adoption of Graviton is now broad enough that AWS cites more than 120,000 customers building on the Graviton family as of the Graviton5 launch. Beyond the aggregate number, individual commitments illustrate scale: Pinterest committed $4 billion in spend to AWS through 2031, and Uber adopted AWS’s Trainium3 AI chips as part of a broader compute deal, both signals of enterprises betting on AWS’s long-term infrastructure roadmap rather than shopping quarter to quarter. Two of the internet’s longest-standing cloud-native companies also anchor opposite sides of this comparison in the public record: Netflix has run the overwhelming majority of its infrastructure on AWS since its well-documented 2008-2016 migration, while Spotify completed a near-total migration to Google Cloud in the mid-2010s and has stayed there since, citing GCP’s data analytics and machine learning tooling as the deciding factor at the time.
Elasticsearch, Databases, and Analytics Workloads
Beyond the headline chip and pricing comparisons, workload-specific data adds useful texture. Google reports Elasticsearch running up to 40% faster on Axion-based C4A instances compared to equivalent x86 configurations, a number that matters directly for any team running self-managed search or logging infrastructure on either cloud. On the AWS side, the GizmoData TPC-H benchmark cited earlier, an 80.475-second Graviton5 result against 89.831 seconds on Graviton4, is one of the clearest independent confirmations that AWS’s own database performance claims hold up under a standardized analytical workload rather than a synthetic microbenchmark. Neither company has published a direct, apples-to-apples benchmark of Graviton5 against Axion N4A on identical workloads as of September 2026, which means any claim that one chip is definitively “faster” than the other for a specific use case should be treated as provisional until you run your own proof of concept against your actual query patterns.
The safest way to read all of this benchmark data together is by workload category rather than by chip. For transactional databases, both platforms show real, independently-measured gains: AWS’s GizmoData-verified TPC-H result and Google’s own Cloud SQL price-performance figures both land in a similar 25-48% improvement band over each provider’s prior generation. For scale-out web and Java workloads, Google’s claimed 85-90% price-performance improvement is the more aggressive number on paper, though it’s currently vendor-reported rather than independently reproduced at the same scale as the AWS TPC-H test. For search and logging infrastructure specifically, Google has the only published third-party-relevant figure (Elasticsearch’s 40% gain on C4A), giving GCP a slight edge in the one benchmark category where AWS hasn’t published a directly comparable number for Graviton5.
6 Scenarios: Which Cloud Fits Your Workload
- Data warehousing and analytics-heavy teams: Google Cloud’s BigQuery, combined with Axion-backed Cloud SQL and the Snowflake/ClickHouse performance data above, makes GCP the stronger default when the primary workload is querying large datasets rather than running general application logic.
- Enterprises with existing compliance and vendor relationships on AWS: AWS’s broader compliance certification catalog and its 120,000+ Graviton customer base mean less friction for regulated industries (finance, healthcare, government) that already have AWS-approved vendor paperwork in place.
- Cost-sensitive, scale-out compute (web serving, batch jobs, stateless microservices): Google Axion N4A’s claimed 90-105% price-performance improvement over comparable x86 instances gives GCP an edge for workloads that scale horizontally and don’t need AWS-specific managed services.
- Teams already using Anthropic’s Claude as their primary LLM: Both Bedrock and Vertex AI host Claude models, but Bedrock’s native integration with the broader AWS ecosystem (IAM, VPC, CloudWatch) tends to reduce plumbing work for AWS-native teams, while Vertex AI’s Claude Fable 5.1 pricing may suit GCP-native teams already billing through Google.
- Cloud-native Kubernetes deployments with no existing lock-in: GKE’s free zonal/Autopilot cluster and Google’s history as Kubernetes’ original developer give it a slight operational edge for greenfield container platforms.
- Global reach and the widest region footprint: AWS’s larger region count and longer operating history still make it the safer choice for multinational deployments requiring in-region data residency across the most countries.
Migration Guide: Moving Workloads Between AWS and Google Cloud
Moving a production workload between AWS and Google Cloud is rarely a weekend project, but the process follows a consistent sequence regardless of direction. Start by inventorying every managed service dependency, not just compute. A workload that looks portable at the VM level often has hidden lock-in through managed databases (RDS vs Cloud SQL), queueing (SQS vs Pub/Sub), or IAM policies that don’t map cleanly to the other platform’s permission model.
- Audit all managed service dependencies (databases, queues, object storage, secrets managers, IAM roles) and map each to its closest equivalent on the target cloud.
- Benchmark your actual workload on both Graviton5 and Axion instances before committing, since vendor price-performance claims are workload-specific and not universally transferable. If you’re staying on AWS, follow a structured Graviton5 migration checklist rather than a straight lift-and-shift.
- Stand up a parallel environment on the target cloud using infrastructure-as-code (Terraform or Pulumi) rather than manual console configuration, so the migration is repeatable and auditable.
- Migrate object storage first using a tool like Google’s Storage Transfer Service (for AWS-to-GCP) or AWS DataSync (for GCP-to-AWS), since storage migration is typically the least risky and easiest to validate independently.
- Re-point DNS and load balancing gradually using weighted routing, moving a small percentage of traffic at a time rather than a hard cutover.
- Migrate stateful databases last, using native replication tools where available, and plan for a maintenance window during the final cutover even if everything else was zero-downtime.
- Re-validate IAM and network security group configurations on the target cloud independently, don’t assume a 1:1 policy translation is safe by default.
- Keep the source environment running in a reduced-capacity standby state for at least one full billing cycle after cutover, in case a rollback is needed.
# Example: pulling an AWS S3 bucket inventory before migrating to Google Cloud Storage
aws s3api list-objects-v2 --bucket my-source-bucket --query "Contents[].{Key: Key, Size: Size}" > inventory.json
# Example: starting a Storage Transfer Service job from AWS S3 to GCS
gcloud transfer jobs create s3://my-source-bucket gs://my-destination-bucket \
--source-creds-file=aws-creds.json \
--project=my-gcp-projectFor teams moving AI workloads specifically, the biggest migration variable isn’t compute, it’s model behavior. A prompt tuned against Amazon Nova Premier or Claude Opus 4.1 on Bedrock will not necessarily produce identical output quality against Gemini 3.1 Pro on Vertex AI, even at similar token pricing. Budget time for prompt re-evaluation and, where possible, run both models against the same evaluation set before switching production traffic.
Pros and Cons: AWS vs Google Cloud
AWS: Pros and Cons
- Pro: Largest global market share (28%) and the broadest region footprint, reducing data residency friction for multinational deployments.
- Pro: 120,000+ customers already running on Graviton, meaning deep community knowledge, tooling, and third-party support for ARM migration.
- Pro: Widest third-party model marketplace on Bedrock, including access to premium frontier models like Claude Opus 4.1.
- Con: Slower year-over-year growth (37% vs Google’s 82%) suggests AWS is investing more in defending existing share than in aggressive new-category price cuts.
- Con: Egress and storage pricing, while competitive, no longer carry a clear cost advantage over Google Cloud at comparable volumes.
- Con: EKS requires more manual node and autoscaling configuration out of the box compared to GKE Autopilot.
Google Cloud: Pros and Cons
- Pro: Fastest-growing hyperscaler by a wide margin (82% YoY), with the newest ARM chip generation (Axion N4A) among the three major clouds as of September 2026.
- Pro: Cheaper baseline object storage ($0.020/GB vs AWS’s $0.023/GB) and a free GKE cluster per billing account.
- Pro: Strong, independently benchmarked real-world results from Snowflake, ClickHouse, and Google’s own Cloud SQL team on Axion infrastructure.
- Con: Still holds only 15% global market share, meaning a smaller talent pool and third-party tooling ecosystem than AWS in absolute terms.
- Con: Premium Tier egress pricing ($0.12/GB) is notably more expensive than AWS’s entry tier ($0.09/GB) unless you opt into Standard Tier routing.
- Con: Axion N4A’s headline price-performance claims are vendor-reported for most workload categories, with limited independent third-party validation published so far.
The Verdict: AWS vs Google Cloud in 2026
There’s no single winner in the AWS vs Google Cloud comparison for 2026, and the data supports treating this as a workload-by-workload decision rather than a platform-wide one. AWS remains the larger, more mature platform by every absolute measure: 28% market share against Google’s 15%, a broader service catalog, more regions, and a larger existing customer base on Graviton. If your organization needs the widest compliance coverage, the deepest third-party model marketplace through Bedrock, or simply has years of AWS-specific tooling already built, switching platforms for marginal price-performance gains rarely justifies the migration cost.
But the growth and product-velocity numbers point the other direction. Google Cloud’s 82% revenue growth against AWS’s 37%, its cheaper baseline storage, its free GKE cluster, and Axion N4A’s independently-benchmarked 13.7-21.7% CPU throughput advantage over comparable instances (per DoiT’s testing) all suggest Google Cloud is the platform iterating faster on cost-per-workload economics right now. For net-new, cloud-native, analytics-heavy, or Kubernetes-first projects with no existing AWS lock-in, that momentum is worth taking seriously. The pragmatic reading of the September 2026 data: default to AWS for scale, breadth, and compliance; default to Google Cloud for new analytics, ARM-native, and Kubernetes-first workloads where price-performance is the deciding factor. Multi-cloud, running specific workloads on whichever platform benchmarks best for that workload, is increasingly the answer most engineering teams are actually landing on rather than picking a single vendor for everything.
Frequently Asked Questions
Is AWS or Google Cloud bigger in 2026?
AWS is significantly bigger by market share, holding 28% of the global cloud infrastructure market in Q2 2026 versus Google Cloud’s 15%, according to Synergy Research Group data. AWS also generates far more absolute quarterly revenue. Google Cloud is growing faster in percentage terms, 82% year-over-year versus AWS’s 37%, but is doing so from a smaller base.
Is AWS Graviton5 or Google Axion faster?
There’s no single definitive answer as of September 2026 because no vendor or independent lab has published a direct Graviton5-vs-Axion N4A benchmark on identical workloads. Graviton5 has the higher raw core count (192 vs Axion’s 72), while independent testing from DoiT found Google’s N4A outperforming AWS’s older M8g (Graviton4-generation) chip by 21.7% in raw CPU throughput. A true Graviton5-vs-N4A comparison would need fresh third-party testing.
Which cloud is cheaper, AWS or Google Cloud?
It depends on the service. Google Cloud Storage’s Standard tier ($0.020/GB-month) is cheaper than AWS S3 Standard ($0.023/GB-month), and Google’s Standard Tier egress ($0.085/GB) undercuts AWS’s entry tier ($0.09/GB). But AWS’s entry-tier internet egress is cheaper than Google’s Premium Tier ($0.12/GB), and compute pricing is difficult to compare directly because the two platforms size their instance families differently.
Should I choose AWS Bedrock or Google Vertex AI for LLM hosting?
Both platforms host comparable models at similar entry-level pricing, with Amazon Nova Premier at $12.50 per million output tokens against Gemini 3.1 Pro at $12 per million output tokens for smaller context windows. Bedrock offers a wider marketplace of third-party frontier models, including Claude Opus 4.1, while Vertex AI’s Gemini Flash tier currently runs cheaper promotional pricing through the end of 2026. The choice usually comes down to which cloud already hosts the rest of your application stack.
Is EKS or GKE better for Kubernetes?
Both charge an identical $0.10 per cluster-hour for standard control planes. GKE includes one free zonal or Autopilot cluster per billing account, giving it a cost edge for single-cluster deployments, and Google’s Autopilot mode automates more node provisioning out of the box. EKS integrates more tightly with existing AWS-native tooling like IAM and Bedrock, which matters more for teams already invested in the AWS ecosystem.
Can I run the same workload on both AWS and Google Cloud simultaneously?
Yes, and it’s an increasingly common approach. Many organizations run analytics and Kubernetes-native workloads on Google Cloud while keeping core transactional and compliance-sensitive systems on AWS. The tradeoff is operational complexity: maintaining infrastructure-as-code, monitoring, and IAM policies across two platforms requires more engineering overhead than a single-cloud strategy, even when it reduces vendor lock-in risk.
How much does it cost to migrate from AWS to Google Cloud?
There’s no fixed number since it depends entirely on workload complexity, data volume, and how many managed services need re-architecting. The largest cost driver is usually engineering time spent re-mapping managed service dependencies (databases, queues, IAM), not the direct data transfer or compute costs. Budgeting for a parallel-run period, keeping both environments live during cutover, is the most reliable way to control migration risk even though it temporarily doubles infrastructure spend.
Related Coverage
- AWS vs Microsoft vs Google Sovereign Cloud: €7.8B [2026]
- AWS ALB vs NLB vs Cloudflare LB: $16 vs $5 Gap [2026]
- AWS Outposts vs Azure Local vs GDC: 5-Yr Lock-In Gap [2026]
- Codespaces vs Gitpod vs Cloud Workstations: $199 Gap [2026]
- AWS Security Hub CSPM Setup: $0.001/Check, 12 Steps [2026]
