AWS cut its rented model's price, but Google still owns the chip and the model underneath it.
On August 3, 2026, AWS dropped Bedrock on-demand prices for OpenAI's GPT-5.6 Luna 80% to $0.20 per million input tokens and $1.20 per million output tokens, closing the raw per-token gap with Gemini Flash, though AWS is still renting that model from OpenAI on largely sold-out Trainium capacity.
We win the largest-scale training and inference workloads on silicon depth.
For AI-native startups, frontier labs and research-heavy enterprises, Google's 8th-gen TPUs (8t/8i) and single-fabric scale to 1M+ chips are a genuine moat: customers run on the same hardware that trains Gemini. AWS Trainium is competitive on cost-per-token but supply-locked and scales across multiple clusters, not one fabric.
External capital is validating TPU demand, not just Google's own books.
On May 19, 2026, Blackstone committed $5B to a Google-backed venture to sell TPU capacity as a service, targeting 500MW by 2027. A third party betting billions on TPU (not Nvidia) compute is independent proof that TPU price-performance is credible at enterprise scale. AWS has no comparable outside vote of confidence in Trainium.
We just closed our biggest historical gap: enterprise delivery muscle.
On June 4, 2026, IBM and Google Cloud launched a joint practice putting thousands of IBM consultants behind industry-specific AI agents (banking, healthcare, telecom, government) on Gemini Enterprise. For regulated buyers who once rejected Google for thin systems-integrator support, there's now a credentialed delivery partner.
Expand section · 2 moreCollapse section▾
AWS added native web grounding to Bedrock, chipping at one of Vertex AI's cleaner edges. On August 4, 2026 AWS made Web Search generally available on Amazon Bedrock for OpenAI models (GPT-5.4, GPT-5.5 and the GPT-5.6 Sol/Terra/Luna variants): a server-side tool that grounds answers in an Amazon-operated web index with zero data egress, turned on by a single parameter in an existing API call at $12 per 1,000 queries. It runs in three US regions (N. Virginia, Ohio, Oregon), covers only OpenAI models on Bedrock (not Claude, Llama, Mistral or Nova) and serves an indexed corpus, with live-web fetch built into the API but not yet enabled. Google's grounding with Google Search still draws on a far larger live index and works across every Gemini model, so the Vertex edge holds for now. What changed is that a buyer who wants grounded answers kept inside their cloud boundary can now get a native version on Bedrock.
Expand section · 13 moreCollapse section▾
Amazon raised its 2026 capital-expenditure guidance to $220B (up from the ~$200B it had held since February), blaming higher memory prices, on its July 30, 2026 Q2 earnings call. Even at $220B, Jassy said AWS still will not have enough capacity to meet all its 2026 demand and expects the same shortfall in 2027. Q2 capex alone hit $54.2B, and free cash flow flipped to a $7.6B outflow.
On July 24, 2026, an AWS network routing failure in its US-WEST-2 (Oregon) region cut internet connectivity for roughly an hour, taking Apple Pay, DoorDash, Reddit, Hulu, PlayStation Network and Twitch offline. AWS traced it to networking devices routing traffic from the region to the Seattle Metro and restored all routes by 4:59 AM PDT, with connectivity within the region itself unaffected. It lands a week after the July 17 Cost Explorer billing bug, a second high-visibility AWS reliability stumble inside a month.
On July 17, 2026, a unit-pricing bug in AWS's billing computation subsystem made Cost Explorer and the Billing Console show wildly inflated estimates (one customer with $0.19 in real charges saw nearly $2.5 billion, others saw figures into the trillions). AWS confirmed the displayed numbers were estimates only: actual usage, invoices and payments were unaffected, and it targeted full corrected data by July 18 noon PT.
AWS is putting its original Bedrock Agents (its flagship agent product, launched November 2023) into maintenance mode, closed to new customers as of July 30, 2026, with a frozen model catalog and no new features. It is one of roughly 20 services AWS moved to maintenance mode in a single June 30, 2026 announcement, alongside Kendra and Q Business. New AWS agent customers now start on AgentCore, which reached general availability only weeks earlier on June 17, 2026.
On June 30, 2026, AWS made Secret Cloud for Industry (ASCI) generally available, letting cleared US defense contractors run contractor-owned classified workloads up to Secret (IL6) directly on AWS, provisioned in days rather than the months an on-prem buildout takes. Northrop Grumman is the first to deploy, and AWS added an up to $20M accelerator to migrate defense-industrial-base, FFRDC and ISV workloads. It opens a classified segment Google Cloud does not serve at this level.
On June 30, 2026, AWS announced the Intelligence Community Accelerated Modernization Framework (ICAMF), committing up to $1B in outcome-based credits through October 2030 for the 18 US intelligence agencies to migrate workloads under AWS's existing C2E contract. CIA Director John Ratcliffe confirmed the CIA will use it. AWS calls it its largest single investment in IC cloud adoption, and it deepens the single-vendor lock-in Google must counter in federal deals.
On June 25, 2026, Amazon committed an additional $13B to expand AWS AI and cloud capacity in India by 2030, taking its planned India AI/cloud investment to more than $21B for 2026-2030 and total India commitments to $48B over the next five years. The new money expands AWS data-center capacity in Mumbai and Hyderabad with Trainium chips and Bedrock. For context, Google pledged $15B in October 2025 for its first Indian AI hub in Andhra Pradesh, and Microsoft pledged $17.5B in December 2025.
AWS brought Graviton5 (its first custom CPU purpose-built for agentic AI) to general availability on June 10, 2026 in the EC2 M9g/M9gd instances, extending AWS's silicon story from training (Trainium) into CPU-bound agentic inference (real-time reasoning, code generation, multi-step orchestration). The 3nm, 192-core chip delivers up to 25% better compute performance than Graviton4, and AWS landed marquee validation: Meta committed to deploy 'tens of millions' of Graviton5 cores under a multibillion-dollar deal, with Uber and Snowflake also onboard, joining 120,000+ existing Graviton customers.
On June 1, 2026, AWS made OpenAI's GPT-5.5, GPT-5.4 and Codex generally available on Amazon Bedrock at OpenAI-matching token rates, its first frontier OpenAI access, landing weeks after Microsoft's exclusivity lapsed. So what for us: it confirms Bedrock is a model marketplace that resells other labs' models. Google still has the only fully-owned frontier model (Gemini) trained on its own silicon.
On April 20, 2026, Amazon agreed to invest up to $25B more in Anthropic (on top of ~$8B prior), and Anthropic committed to spend $100B+ on AWS over ten years, including up to 5GW of Trainium capacity. So what for us: AWS is deepening a dependency on a partner it doesn't own, and Claude still runs on Google Cloud too, so "we need Claude" is not a reason to pick AWS.
In his April 2026 shareholder letter, Andy Jassy disclosed that Trainium2 (roughly 30% better price-performance than comparable GPUs) is "largely sold out" and Trainium3 is "nearly fully-subscribed." So what for us: even buyers who want Trainium can't get it: sell Google Cloud TPU and GPU availability today, not spec-sheet comparisons.
A 13-hour AWS outage in December 2025 was reportedly caused by Amazon's own Kiro agentic AI tool autonomously deciding to delete and recreate an environment (Financial Times, via Engadget). Amazon disputes the framing, calling it "an extremely limited event" affecting only AWS Cost Explorer in one region. So what for us: it's a governance talking point for CISOs weighing AWS for mission-critical AI, but cite it honestly, including Amazon's denial.
In January 2026 AWS quietly raised EC2 Capacity Block prices for its H200 GPU instances ~15% (the p5e.48xlarge jumped from $34.61 to $39.80/hr), reversing its usual "prices only go down" posture amid GPU scarcity. So what for us: this undercuts AWS's claim that its GPU prices only go down, and gives us a price opening.
AWS positions Bedrock as the broadest frontier-model marketplace, now spanning OpenAI (GPT-5.5/5.4, Codex), Anthropic Claude, Meta Llama, Mistral and its own Amazon Nova through one API. This is AWS's own positioning; the strategic tell is that AWS aggregates other labs' models rather than fielding a first-party frontier model of its own.
Expand section · 3 moreCollapse section▾
AWS positions Trainium3 (launched re:Invent 2025) as a cost play, claiming up to 50% savings vs GPU training. This is AWS's own pre-GA benchmark, not independently audited, and the chip's ecosystem still depends on the AWS Neuron SDK, which is narrower than CUDA.
AWS claims the broadest Nvidia GPU lineup of any cloud, pledging to add more than 1 million Blackwell and Rubin GPUs across regions starting in 2026 (its own claim). It's a real strength, but it also underlines how dependent AWS's high-end AI compute is on Nvidia supply.
Google Cloud's differentiator is fabric-scale custom silicon: with TPU 8t and its Jupiter/Virgo networking, Google says it can knit 1M+ TPU chips into a single training cluster: the same infrastructure that trains Gemini, offered to customers. AWS Trainium scales across multiple clusters, not one fabric of comparable reach.
Unlike AWS, Google Cloud applies automatic sustained-use discounts to eligible attached GPUs as monthly usage rises (no upfront commitment required), on top of optional committed-use discounts. This is a structural flexibility advantage for variable AI workloads.
Expand section · 3 moreCollapse section▾
Google's pitch that shifting about 80% of workloads to Gemini 3.5 Flash could save enterprises $1B+ annually still holds, but AWS narrowed the raw per-token gap on August 3, 2026 by cutting Bedrock prices for OpenAI's GPT-5.6 Luna 80% to $0.20 per million input tokens and $1.20 per million output tokens. That price cut is AWS discounting a model it resells from OpenAI, so the number to compare is full workload cost on owned infrastructure, rather than one AWS-set list price.
On August 3, 2026, AWS cut Bedrock on-demand inference prices for OpenAI's GPT-5.6 models, effective July 30: GPT-5.6 Luna dropped 80% to $0.20 per million input tokens and $1.20 per million output tokens, and GPT-5.6 Terra dropped 20%. AWS calls Luna one of the most affordable frontier-class models available.
AWS is raising EC2 Capacity Block reservation prices for its top Nvidia GPU instances about 20% effective July 1, 2026, its second guaranteed-GPU price hike in six months after January's roughly 15% H200 increase. Per-accelerator hourly reservation rates rise to $5.191 for P5, $5.97 for P5e and $14.04 for the Blackwell P6-B300, on the scarce reserved-capacity product enterprises buy for large training runs.
Competitive Battlecard10 across 3 zones›
We win the largest-scale training and inference workloads on silicon depth.
Best for Technical evaluatorFor AI-native startups, frontier labs and research-heavy enterprises, Google's 8th-gen TPUs (8t/8i) and single-fabric scale to 1M+ chips are a genuine moat: customers run on the same hardware that trains Gemini. AWS Trainium is competitive on cost-per-token but supply-locked and scales across multiple clusters, not one fabric.
Expand · 2 moreCollapse▾
External capital is validating TPU demand, not just Google's own books.
Best for Economic buyerOn May 19, 2026, Blackstone committed $5B to a Google-backed venture to sell TPU capacity as a service, targeting 500MW by 2027. A third party betting billions on TPU (not Nvidia) compute is independent proof that TPU price-performance is credible at enterprise scale. AWS has no comparable outside vote of confidence in Trainium.
We just closed our biggest historical gap: enterprise delivery muscle.
Best for Exec / top-downOn June 4, 2026, IBM and Google Cloud launched a joint practice putting thousands of IBM consultants behind industry-specific AI agents (banking, healthcare, telecom, government) on Gemini Enterprise. For regulated buyers who once rejected Google for thin systems-integrator support, there's now a credentialed delivery partner.
Model breadth is now a real fight, not a Google win.
Best for Technical evaluatorBoth platforms offer 200+ models. Bedrock's June addition of OpenAI GPT-5.5 alongside Claude, Llama and Nova is a genuine strength, and its token spend is growing fast. Google counters with first-party Gemini (long-context, multimodal) plus partner models on the Gemini Enterprise Agent Platform. Deals turn on data gravity and price. A bigger model catalog rarely decides them.
Expand · 3 moreCollapse▾
Raw Nvidia GPU availability is a shared constraint, and reserved AWS capacity is getting expensive fast.
Best for Technical evaluatorBoth clouds are supply-limited on top-end Nvidia capacity (TSMC CoWoS and HBM bottlenecks), and both are co-engineering with Nvidia on next-gen Rubin systems. AWS is raising Capacity Block reservation prices a second time in six months: effective July 1, 2026, the per-accelerator hourly rate for reserved P6-B300 hits $14.04, P5e at $5.97 and P5 at $5.191, up roughly 20% from the prior set. A buyer committing to reserved Nvidia capacity on AWS now pays a rising premium for that scarcity. TPU compute gives a buyer a second supply line that reduces both GPU-queue risk and the cost exposure that comes with repeated reservation-price hikes.
Agent-platform maturity is a contested fight. On August 6, 2026, AWS made Bedrock AgentCore runtime instances generally available: AWS-managed EC2 for production agents that can run up to 14 days, with GPU support and several agents collaborating on one host and shared file system. Combined with the production-agent stack AWS shipped at AWS Summit New York (June 17, 2026), AWS now covers long-running, multi-agent and GPU-backed workloads that Vertex AI reps used to pitch as a Google-only strength.
On June 30, 2026 at AWS Summit Washington DC, AWS launched Forward Deployed Engineering (FDE), a $1B organization that embeds thousands of engineers (pods of five to six, working alongside AI agents) inside customers to build and ship production AI in weeks. AWS is the first hyperscaler to stand up this kind of unit; named early customers include the NFL, NBA, Cox Automotive and Southwest Airlines. This directly contests the enterprise-delivery muscle Google claimed via its June 4 IBM joint practice.
AWS wins on incumbency and data gravity. Be honest about it.
Best for Economic buyerWith ~28% market share and revenue nearly double ours, AWS holds the install base, and Jassy's pitch is blunt: customers want inference next to data that already lives in AWS. For a CIO with a decade of AWS footprint, switching cost is the dominant force in the deal.
Expand · 2 moreCollapse▾
If a buyer has standardized on Claude, AWS has the deepest native path.
Best for Eng-led championOver 100,000 customers run Claude on Bedrock, Trainium handles the majority of Bedrock inference traffic, and the AWS and Anthropic co-engineering (Project Rainier, $100B commitment) is hard to match at arm's length. Claude is on Vertex too, but the integration depth on Bedrock is real.
AWS's service breadth and partner ecosystem are a default-platform advantage.
Best for Exec / top-downThousands of services, the largest ISV and SI ecosystem and the broadest compliance coverage make AWS the safe anchor for multi-cloud enterprises. Our $20B revenue and ~14% share mean proportionally fewer integrations and certifications, a gap the rep should acknowledge and route around.
Sentiment5 signal›
Sentiment: enterprise reviewers report unpredictable Bedrock billing: one PeerSpot reviewer cited ~$130 in unexpected charges within two weeks without even deploying a model. Recurring theme: cost opacity as teams move past proof-of-concept.
Sentiment: SageMaker draws consistent complaints about month-end billing shock, a steep learning curve for non-AWS-native teams and "walled garden" lock-in that penalizes multi-cloud strategies (G2/PeerSpot themes).
Sentiment: developers describe Trainium's Neuron SDK as painful outside AWS's happy path ("things tend to fall apart immediately" with custom dependencies) and note the conspicuous absence of public customer endorsements beyond Anthropic. A real adoption-friction signal for the Trainium ecosystem.
Sentiment (the other side): enterprises praise Bedrock for low operational lift and tight IAM/VPC integration within existing AWS security boundaries, a genuine strength for compliance-heavy shops already standardized on AWS. Don't pretend AWS has no fans.
Sentiment (our own weak spot): Vertex AI / Gemini Enterprise Agent Platform draws complaints about complex multi-dimensional pricing and a steep learning curve for teams not already GCP-native. Reps should expect this and have an onboarding/cost-modeling answer ready.
Objection Handling5 objections›
"Google Cloud had that massive 2025 outage. Why trust critical AI to them?"
Raised by Security & regulatedOwn it directly: the June 12, 2025 outage took down 70+ services for hours after an untested global config change, and Google publicly apologized and committed to feature-flagged rollouts and isolation so one fault can't cascade. Every hyperscaler has had a major outage (AWS's us-east-1 included); the question is the fix trajectory, and the market clearly hasn't defected, given 63% Q1 2026 growth.
"Google kills products. You just retired Vertex AI. Why bet our ML platform on that?"
Raised by Exec / top-downThis is grounded and fair to raise. But the April 22, 2026 change is an evolution, not a shutdown: Google explicitly described the Gemini Enterprise Agent Platform as "the evolution of Vertex AI," carrying forward model selection, model building and agent building, with existing workloads preserved. The honest counter is to negotiate continuity into the contract (SLA uplift, migration support, named-service commitments) and point to anchor customers who ran this same analysis and stayed.
"Even Pichai says Google Cloud is compute-constrained. We'll be stuck in a queue."
Raised by Economic buyerTrue, Pichai said it: "We are compute constrained in the near term." Demand is outrunning supply, and Google is spending $180 to 190B in 2026 plus standing up the Blackstone TPU venture to close the gap. AWS admits the same constraint, with Trainium sold out. The right move is to secure committed capacity through the account team now, not to assume an equally-constrained AWS is more available.
"We want Claude, and that means AWS Bedrock."
Raised by Eng-led championNot exclusive: Anthropic itself states Claude is available on all three major clouds, including Google Cloud. Running Claude on Vertex pairs the model with Google's TPU infrastructure and avoids the Bedrock quota waits some customers have hit: same frontier model, your choice of infrastructure and economics.
Cut Log9 removed / revised›
as_of = the date the fact is true as-of · verified_on = when grounding last confirmed the exact wording · is_new = a monitor run touched it <48h ago.| subject_key | section | as_of | verified_on | is_new |
|---|---|---|---|---|
| google-cloud vs aws | ai-infra-growth | q1-2026 | executive_summary | 2026-08-03 | — | false |
| google-cloud vs aws | vertical-integration-edge | current | executive_summary | 2026-04-20 | 2026-06-05 | false |
| google-cloud | tpu-inference-economics | 2026 | executive_summary | 2026-04-22 | 2026-06-05 | false |
| google-cloud | reliability-objection | current | executive_summary | 2025-06-16 | 2026-06-05 | false |
| aws | q-revenue | q1-2026 | snapshot | 2026-04-29 | 2026-06-05 | false |
| google-cloud | q-revenue | q1-2026 | snapshot | 2026-04-29 | 2026-06-05 | false |
| cloud-market | market-share | q1-2026 | snapshot | 2026-04-29 | 2026-06-05 | false |
| aws | ceo | current | snapshot | 2026-04-29 | 2026-06-05 | false |
| amazon | capex-guidance | 2026 | recent_moves | 2026-07-30 | 2026-07-31 | false |
| aws | launch | openai-on-bedrock-2026-06 | recent_moves | 2026-06-01 | 2026-06-05 | false |
| amazon | latest-funding-round | anthropic-2026-04 | recent_moves | 2026-04-20 | 2026-06-05 | false |
| aws | trainium-capacity | 2026 | recent_moves | 2026-04-10 | 2026-06-05 | false |
| aws | security-incident | kiro-outage-2025-12 | recent_moves | 2026-02-21 | 2026-06-05 | false |
| aws | list-price | gpu-capacity-block-hike-2026-01 | recent_moves | 2026-01-05 | 2026-06-05 | false |
| aws | positioning | bedrock-model-marketplace | positioning | 2026-06-01 | 2026-06-05 | false |
| aws | positioning | trainium3-vs-gpu-cost | positioning | 2025-12-02 | 2026-06-05 | false |
| aws | positioning | nvidia-gpu-breadth | positioning | 2026-03-16 | 2026-06-05 | false |
| google-cloud | differentiator | tpu-single-cluster-scale | positioning | 2026-04-22 | 2026-06-05 | false |
| google-cloud | pricing-model | automatic-sustained-use-discount | pricing | 2026-06-05 | 2026-06-05 | false |
| google-cloud | list-price | gemini-flash-token-economics | pricing | 2026-08-03 | — | false |
| battlecard | win | tpu-silicon-scale | battlecard | 2026-04-22 | 2026-06-05 | false |
| battlecard | win | tpu-external-validation-blackstone | battlecard | 2026-05-19 | 2026-06-05 | false |
| battlecard | win | ibm-delivery-partnership | battlecard | 2026-06-04 | 2026-06-05 | false |
| battlecard | contested | model-breadth | battlecard | 2026-06-01 | 2026-06-05 | false |
| battlecard | contested | nvidia-gpu-availability | battlecard | 2026-06-26 | — | false |
| battlecard | lose | aws-incumbency-data-gravity | battlecard | 2026-04-29 | 2026-06-05 | false |
| battlecard | lose | aws-anthropic-claude-depth | battlecard | 2026-03-22 | 2026-06-05 | false |
| battlecard | lose | aws-service-breadth-ecosystem | battlecard | 2026-04-29 | 2026-06-05 | false |
| aws | sentiment | bedrock-billing-surprises | sentiment | 2026-06-05 | 2026-06-05 | false |
| aws | sentiment | sagemaker-complexity-lockin | sentiment | 2026-02-08 | 2026-06-05 | false |
| aws | sentiment | trainium-neuron-sdk-friction | sentiment | 2025-12-15 | 2026-06-05 | false |
| aws | sentiment | bedrock-integration-praise | sentiment | 2026-06-05 | 2026-06-05 | false |
| google-cloud | sentiment | vertex-pricing-complexity | sentiment | 2025-09-30 | 2026-06-05 | false |
| objection | gcp-june-2025-outage | current | objection_handling | 2025-06-16 | 2026-06-05 | false |
| objection | google-deprecates-products | current | objection_handling | 2026-04-22 | 2026-06-05 | false |
| objection | gcp-compute-constrained | current | objection_handling | 2026-04-29 | 2026-06-05 | false |
| objection | claude-only-on-bedrock | current | objection_handling | 2026-04-20 | 2026-06-05 | false |
| objection | aws-market-share-lead | current | objection_handling | 2026-04-29 | 2026-06-05 | false |
| aws | launch | graviton5-2026-06 | recent_moves | 2026-06-10 | 2026-06-11 | false |
| google-cloud | reliability-incident | june-2026 | tracked_facts | 2026-06-15 | 2026-06-16 | false |
| battlecard | contested | agent-platform-maturity | battlecard | 2026-08-06 | 2026-08-15 | false |
| amazon | capex-india-expansion | 2026 | recent_moves | 2026-06-25 | 2026-06-26 | false |
| aws | list-price | capacity-block-gpu-hike-2026-06 | pricing | 2026-06-26 | 2026-06-28 | false |
| battlecard | contested | delivery-deployment-muscle | battlecard | 2026-06-30 | 2026-07-02 | false |
| aws | launch | secret-cloud-for-industry-2026-06 | recent_moves | 2026-06-30 | 2026-07-02 | false |
| aws | launch | icamf-ic-credits-2026-06 | recent_moves | 2026-06-30 | 2026-07-02 | false |
| aws | product-lifecycle | bedrock-agents-classic-maintenance-2026-07 | recent_moves | 2026-07-16 | 2026-07-17 | false |
| aws | reliability-incident | cost-explorer-billing-bug-2026-07 | recent_moves | 2026-07-17 | 2026-07-18 | false |
| aws | reliability-incident | us-west-2-outage-2026-07 | recent_moves | 2026-07-24 | 2026-07-25 | false |
| aws | list-price | bedrock-gpt-price-cut-2026-07 | pricing | 2026-08-03 | 2026-08-04 | false |
| aws | launch | bedrock-web-search-gpt-2026-08 | recent_moves | 2026-08-04 | 2026-08-07 | false |