Amazon’s $200B AI Capex Fuels Rivet‑Driven Hiring Surge

From Visual Tool to Managed Cloud

Rivet launched its managed Rivet Cloud service on October 9, 2025, and months later rewrote its core runtime in Rust, a move that delivered a production-ready AI agent infrastructure and triggered a scramble for systems engineers who can operate it. The infrastructure layer for AI agents had been stuck in a loop: teams either stitched together fragile serverless functions or wrestled with Kubernetes clusters that demanded a platform team just to stay upright. Rivet broke that loop by shipping a single binary that runs the same way on a laptop, a VPS, or a global edge network.

The team open-sourced the core under Apache 2.0 as a TypeScript library for executing graphs: a runtime for durable, stateful processes that survive restarts, network partitions, and version upgrades without losing context. By December 2024 the project added Actors: a primitive that gives every agent, session, or user its own durable process with embedded state, events, and remote procedure calls. The changelog noted the features matter because they run at the edge and remain self-hostable.

Rivet Cloud wraps the open-source Rivet Engine in a control plane that provisions actors in under a minute, routes traffic across a global edge network, and scales automatically without exposing the underlying FoundationDB cluster. The launch emphasized three constraints: no infrastructure setup, no migration of existing cloud accounts, and no vendor lock-in; code stays portable because the runtime is the same binary developers can download and run themselves. Organizations get multi-region deployment, role-based access control, and built-in observability on top of the open-source foundation.

Underneath that managed layer, the runtime underwent a wholesale rewrite. In June 2026 the team released Rivet 2.3, replacing the JavaScript-based RivetKit with a native Rust implementation. State persistence, SQLite storage, and networking now execute in Rust rather than on the Node.js event loop, cutting overhead on the hot paths every actor hits. The same release introduced a reimagined dashboard and fine-grained lifecycle controls for actors. A native Rust SDK followed days later, alongside an Effect SDK for developers who prefer that runtime model.

The new SQLite engine provisions millions of isolated databases that scale to zero, stream changes to frontends in real time, and run at the edge without a separate database server. That technical foundation also enabled agentOS, a lightweight VM for AI agents that cold-starts in roughly six milliseconds and costs a fraction of traditional sandbox environments. Announced in April 2026, agentOS runs WebAssembly and V8 isolates inside a full POSIX environment, giving agents a complete userspace (filesystem, network, process tree) without the overhead of a virtual machine. Durable cron jobs, per-actor queues, and replayable workflows shipped in the months that followed, each building on the same actor primitive.

The progression from open-source library to managed cloud to Rust-native runtime traces a deliberate arc: each layer solves the operational friction the previous layer exposed.

Why Durable Agent Platforms Are Surging

Adoption has compressed from years to months. Enterprises are not just piloting agents; they are moving them into production at a pace that existing serverless and container infrastructure could never absorb.

Token consumption tells the real story. Deloitte surveyed 515 U.S. leaders at enterprises with more than $500 million in annual revenue in December 2025. Today the largest cohort, more than a third, consumes one billion to 10 billion tokens per month on average. Nearly a third already exceeds 10 billion. By 2028, three in five respondents expect to surpass that 10-billion threshold, roughly doubling aggregate token volume in two years. Advanced reasoning models that generate "thinking tokens" before returning an answer drive the spike, along with workloads pushing more than 100 billion tokens per month, a segment Deloitte says will triple from 2026 to 2028.

Public cloud alone can no longer serve that demand cost-effectively. Chris Thomas, U.S. hybrid cloud infrastructure leader at Deloitte, told CIO Dive that enterprises "are creating the workloads that the public cloud alone can't serve cost effectively at this point in time." The same survey found six in seven respondents expect AI infrastructure budgets to increase over the next three years, on average threefold, with large enterprises projecting nearly quadruple current spend. The five largest hyperscalers are collectively planning roughly $700 billion in capital expenditures for 2026 alone, the vast majority directed at AI data centers, GPU clusters, and networking infrastructure.

Yet raw compute is only half the problem. AI inference for enterprise applications — real-time financial modeling, autonomous logistics, live customer interactions — is latency intolerant. It requires 99.999 percent reliability and massive continuous uptime. That reality pushes organizations toward durable execution environments where every agent, session, and user gets a persistent process with its state inside, rather than stateless functions that cold-start on every invocation.

Deloitte's respondents agree: more than half see vendor-specific AI frameworks, half see AI network fabrics, and nearly half see foundational models as levers to increase GPU consumption and hybrid deployment. Regulatory pressure adds urgency. A major technology company's decision to shift a previously open-source model to a proprietary license has reinforced the tilt toward closed solutions. Meanwhile, more than half of Deloitte respondents cite economic uncertainty, nearly half cite organizational challenges and regulatory pressures, and two in five cite talent and skill gaps as factors that could delay AI factory investments, making platforms that reduce operational complexity disproportionately valuable.

Durable AI agent platforms are no longer a niche. They are the layer that holds exploding token volumes, tripling infrastructure budgets, and the hard reliability requirements of production workloads. Rivet's move to a managed Rust-based cloud service sits directly in that current.

The Talent Crunch

The Rust migration itself creates a distinct hiring need. Rewriting RivetKit from JavaScript to Rust for version 2.3 moved state persistence, SQLite, and networking off the JS runtime onto native hot paths. That demands engineers who understand ownership semantics, async runtimes, and systems-level debugging, a profile that overlaps heavily with the infrastructure teams at Cloudflare, Datadog, and Rust-native shops like Materialize or TigerBeetle.

Market data amplifies the pressure. GPU management, Kubernetes, and LLM fine-tuning are listed as the key differentiators.

Rivet Cloud runs on a global edge network but also supports self-hosted single Rust binaries or Docker containers on Postgres, file system, or FoundationDB, plus VPC and air-gapped options. That flexibility sells to regulated enterprises but operationally requires staff who can debug customer environments ranging from Vercel and Railway to bare-metal Kubernetes clusters, a range of environments that scales headcount faster than a pure SaaS model.

How Rivals Are Responding

LangChain enters this moment from a position of measured dominance. The framework reports 200 million monthly open-source downloads, 144,000 GitHub stars, and 7,000 active LangSmith customers, including five Fortune 10 companies. Its abstraction layer spans more than 1,000 integrations across model providers, vector stores, and toolkits, a breadth that makes "vendor lock-in" a marketing target rather than a technical reality.

The clearest signal is LangGraph. LangChain now builds its agents on top of LangGraph's durable runtime, which supplies built-in persistence, checkpointing, rewind, and human-in-the-loop support, capabilities that mirror the actor lifecycle Rivet bakes into its Rust runtime. LangGraph's state-graph pattern handles long-running, asynchronous workflows that traditional serverless functions cannot, and LangSmith Deployment wraps it in a managed platform for "long-running, stateful workflows" with fault-tolerant scaling. That is a direct answer to Rivet Cloud's value proposition: a managed control plane for durable agents.

LangSmith Engine, released in 2026, pushes further into autonomous operations. It clusters production failures into prioritized issues, traces root causes through code and traces, and proposes fixes for review. Fleet adds a no-code layer: "agents for the whole company" that run recurring tasks across daily tools with enterprise security controls. Deep Agents sit above both, offering a batteries-included agent with automatic context compression, a virtual filesystem, and subagent spawning. The stack now spans from low-level orchestration (LangGraph) through observability (LangSmith) to autonomous operation (Engine, Fleet), a full platform play that leaves less room for a standalone runtime like Rivet to differentiate on features alone.

Cloud providers are moving to support AI agents through partnerships, startup investments, and internal projects. Microsoft's agentic workflow push spans GitHub Copilot and Azure AI services. AWS Bedrock Agents and Azure AI Foundry both now offer managed agent runtimes with built-in memory, tool use, and multi-agent orchestration. Google's Vertex AI Agent Builder targets the same no-code enterprise buyer as LangSmith Fleet.

Rivet's Rust rewrite and managed cloud enter this fray with a specific wedge: native performance on the hot paths where actors run. RivetKit's state persistence, SQLite engine, and networking now run in Rust, cutting overhead that compounds at scale. Rivet Cloud's global edge network connects to Vercel, Railway, AWS, and existing VPCs without demanding workload migration, a "bring your own cloud" model that contrasts with the hyperscalers' gravity. The actor model gives each agent a durable process with isolated state, a mental model that maps more cleanly to long-lived infrastructure than graph-based workflows.

The talent market mirrors the standoff. Engineers who ship production agents on different stacks demand different mental models and different operational reflexes. Hiring for "agent infrastructure" now means choosing a paradigm, not just a framework.

Where the Risks Live

Rivet's pitch — durable actors, no cold starts, a single Rust binary you can run anywhere — is the infrastructure team's checklist. But moving mission-critical workloads onto any managed platform introduces risks that don't show up in the getting-started guide. The core tension: Rivet Cloud abstracts away the operational complexity of the Rivet Engine, yet the very durability that makes actors attractive also amplifies the fallout when something goes wrong.

State persistence is the first surface. Rivet Actors store state in SQLite (now a native Rust engine as of the 2.3 rewrite), Postgres, or FoundationDB depending on deployment model. That's a feature for durability, but it means every actor becomes a tiny database with its own consistency guarantees, backup requirements, and corruption modes. Teams accustomed to stateless serverless functions often underestimate the discipline required to run stateful actors at scale: schema migrations, point-in-time recovery, and the "split-brain" scenarios that emerge when actors reconnect after network partitions. The Rust rewrite moved persistence and networking off the JavaScript runtime into native code, which cuts overhead on hot paths but also means bugs in the storage layer are now memory-safety issues in Rust rather than V8 exceptions you can catch and log.

Data sovereignty regulations are already pushing enterprises toward self-hosted infrastructure. The managed service operates on a global edge network and integrates with Vercel, Railway, and AWS; convenient, but each integration adds a jurisdictional boundary. A financial-services team deploying actors that process transaction data in a Rivet Cloud region may find their compliance team blocking the very edge nodes that give Rivet its latency advantage. The self-hosted path (single binary or Docker container) avoids this, but it shifts the burden of hardening, patching, and incident response back to the team that adopted Rivet to reduce operational load.

Multi-agent systems compound the risk. Rivet's actor model assigns every agent a durable, isolated process, but the orchestration layer becomes a centralized control plane. Compromise the orchestration logic or the runtime, and you compromise every agent downstream. The Apache 2.0 license means you can audit the code, but the managed Rivet Cloud control plane is not open source; you're trusting Rivet's operational security for the orchestration tier even if you self-host the actors.

Vendor lock-in claims warrant scrutiny. Rivet's marketing emphasizes "no vendor lock-in" because actors run on Kubernetes, AWS, a VPS, or Rivet Cloud. That's true for the actor runtime. But the graph serialization format and the libraries that execute those graphs are all part of the Rivet ecosystem. Migrating a complex agent graph to another framework means rewriting the orchestration logic, not just moving containers. The switching cost lives in the graph definitions and the team's familiarity with Rivet-specific patterns, a softer lock-in but real for teams that have invested months in them.

The Rust rewrite itself carries maturity risk. That 2.3 release shifted persistence, SQLite, and networking to native Rust. That eliminates a class of runtime overhead, but it also introduces a new codebase surface that hasn't seen the same production hardening as the JavaScript version. Early adopters of the Rust runtime become the stress test. Teams building on Rivet Cloud today are effectively beta-testing the managed service's integration with that new runtime; teams self-hosting the Engine are running the Rust binary directly. Either way, the playbooks for debugging native memory issues, profiling SQLite contention, or diagnosing FoundationDB transaction conflicts remain unwritten.

Finally, there's the talent gap. The hiring surge documented across the AI infrastructure stack means engineers who understand actor models, durable execution, and Rust-based runtimes are scarce and expensive. Zero G Talent's board data shows infrastructure roles at companies like ASML and Stripe commanding wide salary bands. A team that bets on Rivet needs at least one engineer who can read the Rust runtime source, tune the storage backend, and design actor boundaries that don't create cascading failures. Without that expertise, the platform's durability features become liabilities: actors that persist bad state indefinitely, retry loops that amplify load, and failure modes that don't surface in standard observability stacks.

None of this means Rivet is the wrong choice. It means the evaluation criteria for mission-critical agents must answer: can your team operate the storage layer? Does your compliance posture allow the managed control plane? Have you load-tested the Rust runtime's failure modes? And do you have an exit strategy for the graph definitions if Rivet's roadmap diverges from your needs? The platform removes serverless constraints; it doesn't remove the need to think like an infrastructure engineer.

Three-Year Outlook

Analyst projections converge on a steep adoption curve for agent operating systems over the next three years, though the pace differs sharply between enterprise and developer tiers. Deloitte's survey adds a budget signal: AI infrastructure spend is projected to triple for most enterprises by 2028, pulling talent in Kubernetes, GPU management, and LLM fine-tuning into the same labor pool that Rivet and its rivals are fishing.

Concentration at the top coexists with fragmentation below. Multi-agent orchestration is replacing single-agent deployments as the default architecture for 2026. That shift favors frameworks with durable execution guarantees, exactly the space Rivet Cloud targets with its actor model and per-actor state persistence. Governance requirements — data sovereignty, audit trails, air-gapped deployment — are accelerating self-hosted and VPC-hosted adoption.

Three-year outlook: expect the enterprise tier to consolidate around the major platforms plus open-runtime specialists that win on operational maturity: observability, upgrade safety, multi-cloud portability. The developer tier will remain fragmented but standardize on interop standards and observability stacks. Talent demand will track the consolidation — platform integration roles at the top, runtime and infrastructure roles in the specialist layer — with compensation bands already reflecting the split. That same binary — runnable on laptops, VPSes, and a global edge network — is no longer a demo. It's the control plane for the next infrastructure layer.

Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.