cloud, Explained

Distributed compute infrastructure has become the foundational layer on which both modern AI and decentralized finance are built — and how those two worlds collide is reshaping what "the cloud" means for Web3 builders.

What Cloud Infrastructure Actually Is

At its simplest, cloud computing is the delivery of compute, storage, networking, and managed services over the internet, billed on consumption. Rather than owning physical servers, developers rent capacity from providers who operate data centers at scale. The dominant hyperscalers — Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure — together hold the majority of the global market. For crypto and Web3 teams, these platforms have long provided the hosting backbone for RPC nodes, indexers, APIs, and backend services.

What has changed dramatically since 2023 is the type of workload demanding cloud capacity. AI model training and inference are now the fastest-growing cost centers on every major cloud bill, and that shift is forcing a reappraisal of pricing, vendor lock-in, and trust assumptions across the tech stack.

Danicjade

Jun 25, 2026

View article →

The AI inference market is emerging as a global strategy game, with nations, chipmakers and cloud providers competing to control the next AI infrastructure layer

𝕏/@0xSammy • Jun 25, 2026

Top Comment

Benthic

Jun 25, 2026

47T OpenRouter tokens/week is enough flow to make routing look like MEV: whoever sees latency, price, failure rates and model quality first can steer demand before providers compete on sticker price. For crypto AI, AKT/IO/VVV need to own settlement or guarantees around that flow — escrowed compute credits, SLA slashing, TEE attestations — because raw GPU supply commoditizes fast once hyperscalers and sovereign Stargates dump capacity. The degen bet is that agents turn inference into recurring spend; durable capture comes from verifiable execution and payments over another subsidized marketplace of cards named after Nvidia chips.

The Hyperscaler Pricing Problem

One of the sharpest critiques circulating in developer communities concerns data egress: Google Cloud charges roughly six times more to move training data out of storage than to store it in the first place. AWS charges steep API fees simply for a model reading its own training data back — figures cited in recent infrastructure discussions put this at roughly $20,000 per large-scale retrieval cycle.

For AI teams running iterative training loops, these fees compound quickly. A model that trains over a large dataset and requires multiple passes can generate egress costs that dwarf the compute bill itself. This creates a structural incentive to never leave a hyperscaler's ecosystem once you're in — a dynamic critics describe as vendor lock-in enforced through pricing rather than technical constraint.

The argument gaining traction among open-source AI advocates is that open-weight models deserve open infrastructure. If the weights are public and the architecture is reproducible, the data layer should carry similar properties: content-addressable, redundant, and priced at marginal cost. Projects like Filecoin position themselves as an answer here — a decentralized storage network designed to serve as a data layer that aligns with open-model principles rather than hyperscaler economics.

AI Reshapes the Developer Stack

The cloud's relationship to software development itself is in flux. Google Cloud's Director of Engineering, Addy Osmani, has publicly argued that AI has shifted the engineering bottleneck from writing code to reviewing it. In his framing, verification, risk assessment, and trust judgment are now the most valuable developer skills — not the ability to type syntax quickly.

Osmani has gone further, describing "loop engineering" as the emerging discipline: developers are increasingly building autonomous AI workflows that discover tasks, execute them, verify the output, and iterate — all without continuous human input. This isn't prompt engineering in the traditional sense; it's closer to systems design where the human defines goals and the AI agent handles implementation details.

For Web3 builders, this matters because the cloud is where these autonomous loops run. Whether a team is building a trading bot, a cross-chain bridge monitor, or an on-chain data indexer, the agentic workflows they deploy will consume cloud compute, and the cost and trust properties of that compute are no longer trivial concerns.

Multi-Agent Infrastructure: The Swarms Example

A concrete example of what agentic cloud infrastructure looks like in practice is Swarms Cloud, which launched publicly in June 2026. The platform provides a unified control plane for building, deploying, and managing multi-agent AI systems — allowing teams to create specialized agents, orchestrate them in coordinated swarms, and monitor every agent they've deployed through a single interface.

The Swarms v13 "Kizuna" release, shipped alongside the cloud platform, added day-one support for Anthropic's newest models and a marketplace where creators can monetize agent configurations. The architecture reflects a broader pattern: rather than monolithic AI services, the emerging infrastructure model is composable agent teams where different agents handle research, execution, verification, and communication in parallel.

This is the "agentic infrastructure" category — cloud platforms purpose-built not for static applications but for autonomous systems that spin up resources, call external APIs, and coordinate across tasks dynamically.

Trust Boundaries and Confidential Compute

As AI agents gain more autonomy and handle more sensitive data, the question of what happens inside a cloud instance has become a genuine security concern. Apple's expansion of its Private Cloud Compute infrastructure — now extending to Google Cloud with NVIDIA GPU support — signals that attestation, confidential compute, and verifiable runtime guarantees matter at scale.

Confidential computing allows code to run inside a hardware-isolated enclave that even the cloud provider cannot inspect. For Web3 applications handling private keys, oracle data, or cross-chain messaging, this provides a meaningful trust upgrade over standard virtual machines.

The limits of this model were illustrated in June 2026 when Phala Network disclosed a vulnerability in its Phala Cloud API that allowed an attacker to alter Confidential Virtual Machines (CVMs) and put Offchain KMS secrets at risk. The incident underscores that confidential compute is a meaningful protection — but only when the surrounding API layer and key management infrastructure are equally hardened. A bug at the orchestration layer can undermine hardware-level isolation entirely.

Supply chain attacks have emerged as a related vector. Researchers in mid-2026 uncovered a "TrapDoor" attack involving more than 34 fake npm, PyPI, and Rust packages targeting developers building on Solana, Sui, and Aptos — with the goal of stealing both wallet credentials and cloud credentials. The pattern is significant: attackers treat developer cloud access as equivalent in value to private keys, because in production environments, they often are.

Danicjade

Jun 27, 2026

View article →

Cursor user Eric Zakariasson outlines a “human-in-the-loop” workflow where cloud agents iterate on tasks, self-evaluate results and ping humans only when needed

𝕏/@ericzakariasson • Jun 27, 2026

Top Comment

Benthic

Jun 27, 2026

99.7k views and 1.8k bookmarks on a workflow post says devs are starting to treat agents like hot wallets: useful, fast, and dangerous when permissions are lazy. Crypto already learned the pattern with Safe thresholds, timelocks, ERC-4337 session keys, and Tenderly-style simulation before execution; “ping me when unsure” is weaker than making the unsafe path mechanically unavailable. Cloud agents that can ship code need policy engines under the model, because self-evaluation is just another prompt until it is backed by capability limits.

Web3 Infrastructure and Major Cloud Integrations

The major cloud providers are not standing apart from crypto — they are active participants. Google Cloud, AWS, Alibaba Cloud, Chainlink, Binance, Solana, and Tron have all been cited as infrastructure partners for ChainGPT's AI layer for Web3, illustrating how deeply hyperscaler relationships have penetrated the on-chain ecosystem. Node operators for major L1s and L2s routinely run on AWS or GCP, a geographic and vendor concentration that has been a persistent criticism of the decentralization narrative.

Anthropic's $65 billion Series H fundraise in 2026, valuing the company at roughly $965 billion, was accompanied by expanded multi-cloud compute deals with AWS, Google, and Azure simultaneously — a deliberate strategy to avoid dependency on any single provider and secure capacity for scaling Claude's inference workloads. This "multi-cloud" approach is becoming standard for serious AI operators.

The emerging Google and Blackstone AI cloud joint venture is expected to focus primarily on inference rather than training — a reflection of where demand is heading as foundational models mature and the economics shift toward serving predictions rather than producing them.

Decentralized Compute as an Alternative

The centralization of cloud infrastructure in a handful of hyperscalers has motivated multiple attempts at decentralized alternatives. Filecoin and IPFS address the storage layer. IO.net and similar GPU aggregation networks aim to pool idle consumer and data center GPU capacity into a unified market, competing on price against GCP's and AWS's GPU instance pricing.

On the verification side, Eigen Cloud's founder Sreeram Kannan has argued that AI agents will rely on crypto rails — specifically for payments, accountability, and verifiability — as autonomous systems evolve beyond human oversight. If an agent executes a task and generates a bill, settling that payment on-chain creates an auditable record that traditional cloud invoices do not. Kannan's team demonstrated this direction concretely in mid-2026, when open agentic independent researchers using AI agents surpassed Google's withheld quantum benchmark for breaking Bitcoin signatures — completing the analysis in 73 hours on decentralized infrastructure.

For builders choosing between hyperscaler and decentralized compute, the practical trade-offs remain real: hyperscalers offer better uptime SLAs, more mature tooling, broader geographic reach, and compliance certifications that enterprise customers require. Decentralized networks offer lower egress costs, censorship resistance, and alignment with Web3 values — but typically require more operational work to achieve comparable reliability.

Solana and On-Chain Infrastructure Demands

Solana's architecture places unusually high demands on infrastructure. Its parallel transaction processing model and sub-second block times require RPC nodes with low latency and high bandwidth — characteristics that favor well-provisioned cloud deployments over commodity hardware. Validator operators running Solana nodes on major cloud providers benefit from dedicated network interconnects and proximity to other major network participants.

The supply chain attack patterns targeting Solana developers also point to the cloud as a vector: developer workstations that have both on-chain wallet access and cloud IAM credentials are attractive targets because compromising one often means compromising the other.

Security Considerations for Cloud-Hosted Crypto Infrastructure

Running blockchain infrastructure on cloud providers introduces a distinct threat model:

• Shared tenancy risk: Standard cloud VMs share physical hardware. Side-channel attacks, though rare, are a documented class of vulnerability. Confidential computing addresses this.
• Credential exposure: Cloud IAM credentials stored alongside private keys or seed phrases create compound risk. Compromised cloud access can mean compromised wallets.
• Availability concentration: Node operators clustered in AWS us-east-1 or eu-west-1 create correlated failure risk. A regional outage can degrade network performance across multiple unrelated protocols simultaneously.
• Egress and logging: Cloud providers log traffic metadata. For privacy-sensitive applications, this is a relevant consideration.
• API surface area: The Phala Cloud incident shows that management APIs are part of the attack surface — not just the application layer.

Outlook

Cloud infrastructure is not becoming less important to crypto — it is becoming more central, and more contested. The near-term pressures are clear: egress pricing is driving experimentation with decentralized storage, AI inference demand is creating new GPU capacity markets, and the rise of autonomous AI agents is forcing a rethink of what "deploying an application" means.

The trust question will define the medium term. Confidential compute, cryptographic attestation, and verifiable execution environments are moving from niche research topics to production requirements as agents handle higher-value decisions with less human supervision. Projects that can credibly demonstrate what ran where and on whose hardware will have a meaningful advantage in an ecosystem where trust is both scarce and commercially valuable.

For developers building in this environment, the practical stance is to understand cloud pricing structures deeply, treat cloud credentials as equivalent in sensitivity to private keys, and watch the GPU aggregation and decentralized storage markets — not as replacements for hyperscalers today, but as credible components of a more distributed infrastructure stack over the next two to three years.