Modal CEO Erik Bernhardsson’s “new era of software” refers to a shift from AI that only generates answers to AI agents that execute code, use tools and pursue objectives. Those agents need isolated, temporary computing environments—sandboxes—as well as inference, training and reinforcement-learning infrastructure. Modal says it has run more than 3 billion customer sandboxes, but the company’s scale and revenue figures are reported claims rather than independently audited results.
What Bernhardsson means by a “new era of software”
In the interview covered by Bloomberg Technology and MT Newswires on October 14, 2026, Bernhardsson described AI software as moving beyond the chatbot model. A model that answers a question can remain inside an inference service. An agent that builds an application, edits files, calls an API or executes a test needs an operating environment in which those actions can happen safely.
As an Amazon Associate I earn from qualifying purchases.
That makes the central infrastructure problem architectural, not merely a question of obtaining faster GPUs. The platform must run the model, provide tools and files, preserve the right state, enforce limits and isolate generated code from other customers and sensitive systems.
Why AI agents need sandboxes
Execution without unrestricted access
Generated code can be incorrect, destructive or intentionally hostile. A sandbox—typically an isolated virtual machine or comparable environment—gives an agent a bounded place to install dependencies, run programs, inspect outputs and fail without receiving unrestricted access to a company’s production network.
Repeatable workspaces
Agents often need more than one prompt-response cycle. They may create files, run several tools, inspect errors and retry. A useful runtime therefore needs controlled storage, networking rules, process limits and a way to preserve or discard state deliberately.
Auditing and shutdown
Isolation is only one part of security. Operators also need logs, resource limits and the ability to terminate a workload. Those controls make it possible to investigate what an agent did and to stop runaway code or unexpected spending.
What Modal reports about its business
| Reported item | Figure or timing | Qualification |
|---|---|---|
| Funding and valuation | $750 million raised at a $15.75 billion post-money valuation | Reported in the October 14, 2026 Bloomberg Technology/MT Newswires coverage; valuation and funding can change with later transactions. |
| AI sandboxes | More than 3 billion run for customers | Company-reported figure; no independent audit is provided. |
| Sandbox business | About $300 million in run-rate revenue | Reported in an interview rendering; run-rate is not the same as audited revenue. |
| Sandbox launch | 2023 | Timing described by Bernhardsson’s account of the product’s history. |
Bernhardsson’s account says demand accelerated after models became capable of generating fuller applications. The reported numbers indicate substantial activity, but they do not establish profitability or the quality of the underlying workloads. Available coverage does not disclose audited gross margins, customer concentration, retention or independently verified economics.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteHow this fits the three waves of AI software
Microsoft CVP Tim Bozarth describes a progression from questions to tasks and then to coordinated objectives:
“The first wave is asking an LLM to answer questions. The second wave is deploying agents to perform tasks and automate workflows. We’ve now entered a third wave, where success comes from managing a network of agents aligned to complete an objective.”
In the third-wave model, code is an output rather than the primary input. A person or business specifies an objective; a network of agents plans, writes code, invokes tools, evaluates results and revises its approach. That increases demand for runtime controls because the system’s behavior unfolds over many actions instead of one response.
NVIDIA CEO Jensen Huang calls the broader change a “platform shift,” comparing it with earlier internet and cloud transitions that created new application layers. Modal’s argument is that agent runtimes and secure execution environments are part of that new layer.
The infrastructure stack behind agent workloads
| Layer | Role in an agent system | Questions buyers should ask |
|---|---|---|
| Inference | Runs the selected model and produces decisions or code. | What are the price, latency, throughput and model-choice trade-offs? |
| Sandbox or virtual machine | Executes generated code and tools in an isolated environment. | Can workloads be contained, inspected, limited and stopped? |
| Agent runtime | Manages files, networking, tools, persistent state and multi-step execution. | Is state persistent when needed, and disposable when it should be? |
| Training and reinforcement learning | Improves models and agent behavior through fine-tuning, evaluation and feedback loops. | Can the same platform run experiments and measure whether behavior improves? |
| Developer workflow | Connects APIs, observability, reproducibility and coding-agent workflows. | Can teams reproduce a run and diagnose an agent’s actions? |
| Capacity and portability | Determines access to GPUs, regions and alternative environments. | How easily can workloads move if capacity, geography or pricing changes? |
Security is the practical dividing line
For an agent platform, security is not limited to model filtering. Teams must define what the runtime can read, write, install, contact and retain. A sandbox that is fast but cannot provide meaningful network controls, logging or termination may be unsuitable for production code execution. Conversely, strict isolation can add startup time, compatibility work and operational cost.
Buyers should test the complete path: launch an environment, provide only the required credentials, run untrusted code, observe outbound connections, collect logs, enforce resource limits and destroy the environment. They should also establish how secrets are injected and revoked, how files are scanned, and whether a failed run leaves data behind.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the commercial claims do—and do not—show
Modal’s reported funding, sandbox volume and run-rate revenue show that investors and customers are assigning significant value to AI infrastructure. They do not prove that every sandbox is a paid production workload, that usage is evenly distributed across customers, or that the business is profitable. “Run-rate revenue” annualizes a current pace; it is not a substitute for audited recognized revenue.
McKinsey reports that CIOs, CTOs and chief data and analytics officers lead nearly 60 percent of AI platform buying decisions. That makes enterprise requirements—governance, observability, predictable capacity and security—central to the market, not secondary features for developers.
What engineering and technology leaders should evaluate
- Start with the workload. Separate simple inference from agents that compile code, browse internal data, run tests or operate continuously.
- Map permissions. Define the minimum file, network, tool and credential access each agent needs.
- Measure end-to-end economics. Include model calls, sandbox startup, storage, GPU time, retries, observability and human review.
- Test failure behavior. Verify timeouts, termination, cleanup, rate limits and recovery after a tool or model error.
- Check portability. Confirm whether images, files, logs and agent state can be exported or reproduced elsewhere.
- Set evaluation loops. Track task success, unsafe actions, latency and cost before expanding autonomy.
Why the “new era” claim matters
The phrase is best understood as a description of software architecture rather than a claim that every application will become autonomous. As agents take on longer tasks, the boundary between a model API and a cloud operating environment becomes less clear. The winning platforms will have to make execution safe, observable and economically predictable while still giving developers access to capable models and sufficient compute.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
