ComputeSDK releases 2026 100k Scale Invitational results across 6 sandbox providers

• ComputeSDK's kickoff thread says the 2026 100k Scale Invitational measured peak concurrency at 100,000 sandboxes, with Namespace powering the benchmarks.
• According to ComputeSDK's methodology note, providers went through multiple test runs over several weeks, and its load-switching description framed the exercise as a stand-in for a customer arriving with massive load.
• ComputeSDK's participant list named six providers that completed the run: Modal, Tensorlake, Northflank, Declaw AI, E2B, and Isorun.
• The provider-specific posts for Modal, Tensorlake, Northflank, Declaw AI, E2B, and Isorun each point to a dedicated results page rather than a single vendor-picked screenshot.
• E2B's completion post shipped with a status card showing 100,000-plus peak concurrent reached in 29 seconds on 2 CPU sandboxes, and ComputeSDK's follow-up immediately turned the thread toward a bigger target: 1 million sandboxes.

ComputeSDK did not just post a winner graphic. It posted a benchmark frame, a weeks-long tuning process, and a set of provider-specific result pages you can open from the master results link. The weirdly useful bit is how explicit the thread's setup post is about simulating a customer switching huge load, while the E2B card gives at least one concrete timing datapoint, and the 1 million sandboxes post shows where the organizers want to push next.

The public framing is simple: hit 100,000 peak concurrent sandboxes. The more interesting part is the test process around it.

ComputeSDK wrote that each provider went through several runs over multiple weeks, with the benchmark stretching provider infrastructure in collaboration with the organizer via its methodology note. In a separate post, ComputeSDK's load-switching description said the exercise was meant to resemble a customer suddenly bringing massive load to a provider.

That gives the invitational a slightly different shape than a one-off stress test. The story here is repeated infra tuning under a fixed public target, not a single heroic demo.

The completion set is short and concrete:

• Modal, per ComputeSDK's Modal thread
• Tensorlake, per ComputeSDK's Tensorlake thread
• Northflank, per ComputeSDK's Northflank thread
• Declaw AI, per ComputeSDK's Declaw AI thread
• E2B, per ComputeSDK's E2B thread
• Isorun, per ComputeSDK's Isorun thread

Each of those posts is phrased as a successful completion, which makes this release read more like a scored graduation board than a ranked leaderboard. The tweets in evidence do not publish a comparative table, only per-provider result links.

E2B is the only provider in the evidence bundle with OCR-visible numbers attached to the tweet card. The image shows:

• Status: Complete
• Target reached: 100,000+ / 100,000 peak concurrent
• Time to target: 29s
• Sandbox CPU: 2 CPU

Because the E2B post is a standalone embed, the attached card carries most of the value without extra narration. It is still the clearest glimpse in the evidence of what a provider result artifact looks like.

The release is distributed across individual result pages, not collapsed into one announcement thread. ComputeSDK posted a master link in its results post, then separate "full results" links for Tensorlake, Northflank, Declaw AI, E2B, and Isorun. Modal's thread also points readers to a dedicated page through its thread link.

That structure matters for one practical reason: it suggests the benchmark is being published as provider-by-provider artifacts, which is a better fit for engineers comparing run details than a single marketing recap.

The benchmark thread did not end at 100,000. Later posts from ComputeSDK said, "Oh, we're in for 1 million sandboxes!" in one follow-up and "More requests for 1m sandboxes!" in another.

That is the cleanest new fact at the end of the story: the invitational results landed as proof of a public 100k bar, and the immediate response from the organizer was to float a 10x next milestone.