Cloud Strategic Report — Week May 9
Central Idea
Multicloud stopped being an aspirational strategy this week: with AWS Interconnect GA, the May free tier live, and 87% of enterprise already in multi-cloud, the question is no longer whether your architecture is multicloud — it's whether your multicloud networking is private, fast, and zero-cost.
Executive Conclusions
- AWS Interconnect GA + May free tier marks the inflection point where private multicloud transitions from differentiator to commodity (🟢 High conviction) — The no-egress-fee pricing model with 500Mbps free per region eliminates the last economic barrier to private connectivity between AWS and Google Cloud.
- The $100B Trainium deal between AWS and Anthropic redefines the cloud AI business model: proprietary silicon as durable competitive advantage, not just compute price (🟢 High conviction) — AWS is not selling GPUs at market price; it is betting that proprietary silicon (Trainium2/3) creates a performance-per-dollar moat that no competitor can match with standard hardware.
- Google Cloud +63% YoY in Q1 2026 suggests that AI-native infrastructure advantage is translating into real market share capture, not just benchmarks (🟡 Medium conviction) — The growth is real, but sustainability depends on whether the TPU and proprietary model differentiation holds as AWS scales Trainium.
Week-to-Week Comparison
Compared to May 2, the signal that moved most was the activation of the AWS Interconnect free tier for May 2026, converting multicloud private networking from a priced service into a zero-marginal-cost baseline. What sustained its direction was the confirmation that Azure and OCI integrations are formally on the roadmap.
Continuity: Accelerates the multicloud-as-standard pattern that became visible in late April when AWS announced GA of Interconnect with Google Cloud — this week's free tier converts the trend into favorable economics, eliminating the last argument against private multicloud as an architectural default.
01. Key Changes and Drivers
Market Signals
- AWS Interconnect GA — free tier active in May: AWS announced GA of AWS Interconnect multicloud on April 14, with Google Cloud as the first partner, and activated the free tier in May: one 500Mbps connection free per region across five available region pairs (US East/West/Oregon, Europe London/Frankfurt). The subsequent pricing model charges for bandwidth without per-GB egress fees. AWS published the underlying spec on GitHub under Apache 2.0, signaling they are betting on establishing a de facto standard for multicloud connectivity.
- 94% of enterprise uses cloud; 87% adopts multi-cloud: The average enterprise uses 4.8 distinct cloud providers. This is not aspirational — it is the current state. The relevant architectural question is no longer "multi-cloud?" but "how do we manage networking between providers without it becoming a failure point or a variable cost?"
- Google Cloud +63% YoY in Q1 2026; AWS +28%; Azure +40%: Google Cloud is the hyperscaler with the highest relative growth, driven by its position in AI-native infrastructure. AWS remains the largest in absolute value. Azure grows through integration with Microsoft 365 and GitHub Copilot in enterprise.
- AWS + Anthropic: $100B for 5 gigawatts of Trainium capacity through 2036: The deal covers Trainium2 and Trainium3, with a target of 1GW deployed before end of 2026. This is not a software agreement — it is a proprietary silicon commitment at a scale that requires manufacturing, supply chain management, and data center operations that only AWS can sustain.
Launches and Integrations
- Anthropic doubled Claude Code rate limits via SpaceX Colossus One: +300MW of new capacity, equivalent to 220,000 NVIDIA GPUs. The architectural implication: frontier AI providers are building their own compute backbone independent of hyperscalers when hyperscalers cannot scale fast enough.
- AWS + Anthropic: Claude Opus 4.7 on Amazon Bedrock (previous week, Apr 20): The integration cycle between frontier model and cloud platform has compressed significantly — the gap between model launch and Bedrock availability is now weeks, not months.
- MCP as universal integration layer: With 97M installs, MCP is the de facto protocol for connecting agents to external APIs and data. This has direct cloud networking implications: secure connections between agents and external data sources will increasingly travel over private backbones like Interconnect.
Market and Competition
- Google Cloud vs. AWS in the AI infrastructure race: Google holds advantages in custom silicon (TPUs), proprietary AI models (Gemini), and platform (Vertex AI), which together form the most complete argument for a fully proprietary AI stack. AWS counters with Trainium, Bedrock as a multi-model platform, and the largest global network scale. Azure competes through integration with the Microsoft ecosystem (365, GitHub, Teams) more than through native AI infrastructure.
- Vertical integration as strategy: All three hyperscalers are racing to control silicon → model → platform → application in a vertical stack. The one that achieves fluid integration across all four layers has the most defensible moat. This week, AWS took the largest step in silicon (Trainium deal), while Google maintains the advantage at both ends (TPUs + models).
02. Winners and Losers
Winners
- AWS (infrastructure play): The $100B deal with Anthropic and the Interconnect free tier give AWS two simultaneous advantages: the most powerful AI provider embedded in its platform, and the de facto standard for multicloud connectivity. If Azure and OCI join Interconnect (confirmed in roadmap), AWS will have defined the protocol the entire market depends on.
- Google Cloud (growth + AI stack): +63% YoY is not accidental — it is the materialization of the AI-native infrastructure bet. Google Cloud is the first AWS Interconnect partner, giving its customers private connectivity with the leading hyperscaler without sacrificing their position.
- Anthropic (infrastructure independence): The combination of Colossus One (SpaceX) + Bedrock (AWS) + Vertex AI (Google) + Foundation Model API means Anthropic has access to multiple compute and distribution paths. It is the AI provider with the highest infrastructure diversification in the sector.
- Enterprises on strategic multi-cloud: The Interconnect free tier eliminates the economic friction of testing private connectivity between AWS and Google Cloud. Companies that already had workloads in both can now connect them at zero variable cost.
Losers
- Third-party multicloud networking vendors: Products like Aviatrix, Megaport, or SD-WAN services that lived off the connectivity gap between clouds now face the problem they solved being converted into a free hyperscaler feature.
- On-premise data centers without a cloud-adjacent strategy: With 94% of enterprise in cloud and 87% in multi-cloud, the pure on-premise segment is increasingly on the defensive. Investment in Trainium and TPUs makes cloud AI compute cost increasingly difficult to match on-premise.
- AI providers dependent on a single hyperscaler: With Interconnect GA, multi-cloud architecture becomes easier. AI companies on only AWS or only Google Cloud have greater dependency risk than before.
03. Incentives and Differentiation
Core incentive structure: Hyperscalers no longer compete solely on compute price — they compete to be the "operating environment" where AI performs best. AWS bets that proprietary Trainium + multi-model Bedrock is the best production AI stack. Google bets that TPUs + Vertex AI + Gemini is the most integrated stack. Azure bets that Microsoft 365 and GitHub Copilot integration is the natural entry point for enterprise. In this context, whoever controls proprietary silicon holds the greatest long-term leverage.
Zones of real differentiation: Proprietary silicon (Trainium vs. TPUs) is real and sustainable for years — it cannot be replicated in 12–18 months. Vertical model-platform integration (Bedrock, Vertex AI) is a real differentiator today. Private multicloud connectivity no longer differentiates — Interconnect GA democratized it.
Ongoing commoditization: Basic inter-cloud connectivity, standard GPU compute (H100 no longer differentiates), and distributed storage services remain fungible and price-competitive. The differentiation frontier has moved up: specialized silicon, AI-native integration, and multi-agent governance are the new competitive terrain.
04. Bottlenecks
- Inter-cloud latency in latency-sensitive workloads: Interconnect delivers high speed and privacy, but the physical latency between different providers' data centers remains a constraint for sub-10ms workloads. The free tier does not resolve physics.
- Identity and access management in multi-cloud environments: With 4.8 providers per enterprise on average, multi-cloud IAM is the most urgent unsolved problem. No universal standard exists; every provider has its own implementation.
- Engineer shortage with multi-cloud expertise: The skill gap for operating Kubernetes + networking + security + cost optimization across multiple clouds simultaneously is real and widening faster than the training curve.
- Cost visibility in multi-cloud: The egress fee model is complex per provider. Interconnect eliminates the problem between AWS and Google, but total cost visibility in environments with 4+ providers remains problematic without dedicated FinOps tooling.
05. Architecture Impact
What architects need to incorporate into their decisions:
- Redesign the inter-cloud networking model: With Interconnect GA and free tier, there is no economic justification for using public internet or VPN to connect critical workloads between AWS and Google Cloud. Migrate existing connections to the Interconnect model before the free tier expires or changes its pricing.
- Plan for Azure in Interconnect: Azure is confirmed on the AWS Interconnect roadmap. Architects designing networking today should leave extension space toward a third provider without requiring a redesign.
- Evaluate proprietary silicon for AI-intensive workloads: For training and inference of large models, Trainium vs. standard H100 cost may be 30–50% lower on workloads AWS has optimized. Run the benchmark before the next scaling cycle.
- Implement multi-cloud FinOps from day one: With 4.8 providers on average, the first source of optimization is cost visibility. Multi-cloud FinOps tools now integrate with Interconnect metrics — this simplifies cost attribution per data stream.
06. Suggested Decisions
- Activate AWS Interconnect with Google Cloud this month — The free tier is available and setup is straightforward. The cost of not doing so is continuing to use public internet or VPN for traffic that could be private and free of variable latency.
- Include Trainium in the compute benchmark for upcoming AI workloads — If you use AWS Bedrock, the underlying silicon starts to matter. Request specific Trainium2 vs. H100 benchmarks for your workloads before the next reserved capacity commitment.
- Establish a centralized multi-cloud IAM — With the growing provider count, fragmented IAM is the largest security risk. Evaluate solutions like HashiCorp Vault, AWS IAM Identity Center with federation, or Google Cloud IAM with cross-cloud federation.
- Map current egress fee dependencies — With Interconnect, egress between AWS and Google Cloud carries no per-GB cost. Audit which current data flows could migrate to Interconnect and quantify the savings before the free tier model changes.
07. Risks
| Risk | Severity | Mitigation |
|---|---|---|
| Interconnect free tier is temporary; pricing model may change when Azure integrates | Medium | Design so that private networking is the default, not dependent on free tier for the business case |
| Trainium lock-in: workloads optimized for Trainium are difficult to port to H100 or TPUs | High | Maintain an abstraction layer (frameworks like JAX, PyTorch) above the silicon; do not optimize at hardware level without an exit plan |
| Multi-cloud skill gap widens faster than internal training capacity | Medium | Identify the 2–3 most critical engineers for multi-cloud operations and dedicate training time now, before the next expansion |
| Azure without Interconnect integration creates asymmetry in environments already using all three hyperscalers | Medium | Plan the integration as upcoming: document which workloads you would move to Interconnect as soon as Azure is available |
08. Weak Signals
- 🟢 The AWS Interconnect Apache 2.0 spec could attract regional providers: If second-tier cloud providers (OVH, Hetzner, Linode) implement the protocol, private multicloud extends beyond the three hyperscalers. Regional adoption would be the first indicator in 60–90 days.
- 🟡 Colossus One (SpaceX) as a third AI infrastructure pillar: Anthropic's partnership with SpaceX for 300MW of additional capacity suggests that frontier labs are building compute options outside the hyperscaler ecosystem. If this scales, it could create a distinct "native AI cloud" market separate from AWS/Google/Azure.
- 🟡 Multi-cloud FinOps as a differentiated investment category: With 87% of enterprise in multi-cloud and 4.8 providers on average, cost complexity is a real problem. FinOps tools that integrate Interconnect metrics from day one could become critical infrastructure in the operational stack.
Open Question
Open question for next week: Will Azure announce an integration date with AWS Interconnect before Google Cloud Next (June 2026), or will it wait until it has its own multicloud networking product before joining? Azure's decision will reveal whether the strategy is to join AWS's standard or compete against it.