The AI build-out filed its paperwork this week. A data center operator that was mining bitcoin three years ago put a number on what it costs to carry a $103B contract book, a Chinese lab put a 600B-parameter model behind an API priced at one-eighth of a frontier competitor, and a chipmaker and a startup both attacked the same constraint from opposite ends: intelligence still needs a building, and it still needs to fit in memory.
Below: seven stories from September 18 to 21, 2026.
1. Nscale files for a New York listing and a valuation of up to $35B
Nscale submitted an S-1 registration statement to the SEC on September 18 and applied to list its ordinary shares on the New York Stock Exchange under the ticker NSCL. Goldman Sachs, JPMorgan and Morgan Stanley lead the underwriting, the offering size and price range are unset, and the Financial Times reports the company is seeking a valuation of up to $35 billion. It also plans to open a retail subscription channel to UK investors through RetailBook, an unusual move for a US listing by a London-headquartered issuer.
The filing is the first full look at the economics. For the six months to June 30, 2026, Nscale reported revenue of $140.6 million against a net loss of $1.02 billion; a year earlier the same period produced $10.4 million of revenue and a $368.9 million loss, so the top line grew roughly 1,252%. The company states total contracted value above $103 billion, up from $100 million two and a half years ago, and the largest single commitment is Anthropic's agreement last month to pay $45 billion to rent capacity from the West Virginia campus. As of the end of August, 7 megawatts ran in its own data centers and 48 more in leased facilities, carrying about $2.6 billion of contract value, with roughly 1.3 gigawatts still in planning or construction.
Nvidia appears on the other side of the table in four roles: chip supplier, investor with more than $2 billion committed, computing customer under leases worth about $1.2 billion, and credit support, including a guarantee of up to $860 million of lease obligations for a Texas facility. The filing also shows Nvidia taking $1 billion of a $3.1 billion convertible bond issue agreed this month. That circularity is disclosed, not hidden, and it sits next to the three risks any buyer has to price: a pipeline concentrated in a handful of AI labs, capital intensity that requires continuous financing, and grid and memory supply that are both tight. The company was spun out of Australian bitcoin miner Arkon Energy in 2024, raised $2 billion at a $14.6 billion valuation in March from Nvidia, Dell and Nokia, and has accumulated about $3.7 billion of equity and more than $5 billion of debt.
— Nscale · Financial Times
2. StepFun ships Step 5 Preview and opens a 600B MoE at one-eighth of Opus 5 pricing
StepFun released Step 5 Preview on September 20, a flagship base model built for agentic work in coding, software engineering, professional knowledge work and finance. The architecture is a sparse mixture of experts with 600 billion total parameters and 27 billion active per token, a 1-million-token context window, and native text and vision input. The API is open in full today, and the company will release the model weights under an open license on October 15.
The claims are anchored to cost. StepFun reports a score of 44 on the Artificial Analysis Intelligence Index, placing it in the top three open models worldwide, with a per-task cost of one-eighth of Claude Opus 5. On the CLI subset of Agents' Last Exam, the FrontierFinance investment-research benchmark and the DRACO deep-research suite, the model trails only GPT-6 Astra or Claude Opus 5 and leads every other open entrant. The company also built its own StepCodeBench spanning 553 repositories, 9 task categories and 33 programming languages to test success rate and stability across scenarios rather than on a single domain.
The long-horizon numbers are the ones worth watching. On a 24-hour GPU kernel optimization task, Step 5 Preview pushed an MLA kernel to 508 TFLOPS peak, ahead of the 493 TFLOPS the company measured for Claude Opus 5, and in an automated post-training experiment it lifted Qwen3-30B-A3B accuracy on AIME24 from 53.3% to 60.0%. StepFun also demonstrated continuous execution beyond three hours on an ESP32 board rebuild and end-to-end financial research runs. The release lands in a market where Chinese open-weight vendors now compete on delivered work per dollar, and a 600B-class model with open weights arriving two weeks from now will give anyone running their own inference a new reference point.
— StepFun · ifeng Tech
🔗 StepFun · ifeng Tech
3. The FAA starts flying AI over Washington this week
The FAA signed a $875 million, 12-year contract for SMART, an AI airspace management system built by Air Space Intelligence, and the pilot goes live on September 21 at three Washington-area airports: Reagan National, Dulles International and Baltimore/Washington International. The agency targets nationwide rollout by the end of 2028. Conventional air traffic control works on a decision window of roughly 15 minutes; SMART extends conflict prediction to two hours, which moves the operation from reacting to traffic as it develops toward planning around it before it arrives.
The vendor is a startup, not an incumbent. Air Space Intelligence's Flyways AI platform has run for more than five years at Alaska Airlines and other carriers, and the company beat Palantir and Thales for the contract. The win says something about how these procurements are being judged: five years of operational flight data at named airlines counted for more than a defense prime's certifications.
The context is a workforce problem the FAA has stopped pretending it can hire its way out of. The agency cut its 2026 target for certified professional controllers from 14,633 to 12,563, and facility aging continues to constrain throughput. A system that gives controllers two hours of lookahead instead of fifteen minutes changes what a shortage of humans actually costs, because the scarce resource stops being reaction speed and becomes judgment over a longer horizon. That is also the risk: air traffic control is the highest-consequence deployment of a prediction system in civilian infrastructure so far, and the two-year pilot window exists because nobody knows how the failure modes behave at scale.
— FAA · Air Space Intelligence
🔗 FAA · Air Space Intelligence
4. Huawei stacks a data center four layers high
Huawei published what it calls the industry's first three-dimensional data center design, a four-layer vertical stacking architecture aimed at the AI clusters and gigawatt-scale campuses now pushing past what a flat footprint can absorb. Land and grid connections have become the binding constraint on AI build-out on both sides of the Pacific, and vertical construction is the direct answer to a site that cannot get wider.
The design separates the three resources that fight each other in a conventional hall. Modular stacking and zoned cooling let compute, power delivery and heat rejection be configured independently of one another, which matters because accelerator density has been rising faster than either power distribution or cooling can follow. The scheme also supports phased expansion, so a campus can bring up capacity in stages as tenants sign, and it contains local failures so a fault in one zone does not propagate through the building.
Huawei's announcement follows its Ascend 960 supernode roadmap and the company's push to sell complete AI infrastructure rather than chips alone. A reference design for stacking matters mostly if it shortens the path from a signed power agreement to tokens produced, and Huawei is betting that in markets where land is scarce and grids are queued, the vendor who owns the building pattern owns the deal.
— Huawei
🔗 Huawei
5. A ternary 27B model fits agentic tool calling into 6GB
PrismML released Ternary Bonsai 2 27B, a Qwen3.8 27B rebuilt with end-to-end ternary weights plus FP16 group-wise scaling. The result is 1.76 effective bits per weight, roughly a ninth of the original footprint, and 5.9GB on disk. The company reports 98.2% aggregate benchmark retention and a score of 77.57 against the base model's 79.74 on agentic tool calling, and it ships GGUF files plus a custom llama.cpp fork. Under 6GB, the model runs in a browser through WebGPU.
What separates this release from the usual quantization claim is the verification. Vendor-published numbers in this genre usually stand alone, but independent community runs appeared the same week: a head-to-head against a Qwen3.8 27B IQ3_XXS build at 10.18GiB, and a separate report that the PQ2_0 quantization does not collapse into unusability. The size claim looks real. The quality claim still rests on benchmarks chosen by the vendor, and a LocalLLaMA thread titled "Ternary Bonsai is a headless chicken" documents erratic generation outside the eval set, which is the classic failure mode of aggressive quantization.
The deployment consequence is concrete. Agentic-class tool calling inside a 6GB envelope means the model fits on hardware you can place inside a compliance boundary, which removes the main argument for sending clinical, legal or financial context to a hosted frontier endpoint. It also lands in the same week Huawei and DeepSeek both attacked the memory budget from the systems side, and the pattern across the three is the same: the memory footprint is now an architectural input, not a deployment afterthought.
— PrismML · The AI Wire
🔗 PrismML · The AI Wire
6. xAI closes its Galaxy event by building a company with Grok Bot in 72 hours
xAI ended its three-day Galaxy event in San Francisco on September 17 with a demonstration: three employees, Lauren Tan, Matt Palmer and Roshan Sadanani, spent 72 hours building and operating a functioning company with Grok Bot as the workforce. The persistent agent ran on its own virtual computer with its own tools and applications, and the showcase doubled as the launch platform for three enterprise products: the Voice Agent API, the Voice Agent Builder and the Agent Tools API.
The pricing is the aggressive part. The Voice Agent API opens at $0.08 per minute for speech-to-speech processing, below what Deepgram and ElevenLabs charge, and the no-code Voice Agent Builder has been in beta since July 1 for teams that want production voice agents without engineering staff. On the consumer side, xAI is rolling out a hands-free voice mode on desktop and mobile powered by Grok Voice Think Fast 2.0, which the company says runs 1.4 times faster than its predecessor and transcribes 1.5 to 2 times more accurately than Deepgram Nova 3 and ElevenLabs Scribe v2.
The enterprise pitch is containment plus proof. Every agent runs in its own virtual machine under a VM-per-Agent architecture backed by a $50 million investment in AIR Security, and the platform ships access, network and audit controls, Action Recording and OpenTelemetry Export for compliance teams. xAI's own Haggle Bot, a procurement agent launched September 3, has surfaced more than $100,000 in savings by monitoring vendor spend and contracts, and Grok and Cursor Enterprise customers get the whole platform free for two weeks. The interface is built around five objects: Bots, Chats, Prompts, Tools and Artifacts, with prompts saved as Skills or triggered automatically as Routines. Whether VM-per-Agent scales to thousands of concurrent corporate agents at acceptable cost is the open question; the isolation guarantee is easy to state and expensive to run.
— xAI · Forkast
7. Faraday Future puts nine robots on sale at once, starting at $89,900
Faraday Future closed its "919" launch event in Gardena, California on September 19 by putting nine robots on the market simultaneously and opening retail sales of the All-New Futurist humanoid at $89,900, including a $10,000 Skills package of pre-loaded task capabilities. The lineup spans three body types at three size classes: the Futurist, Master and Master Mini humanoids, the Aegis and Navi quadrupeds, and Faber, a wheeled mobile manipulator with dual arms for warehouse work.
The flagship carries 28 motors with peak torque up to 500 newton-meters, a screen-based face that interacts in roughly 50 languages, and, by the company's claim, is the first full-size humanoid sold in the United States to run Nvidia's Sonic full-body motion control system natively. Four industry packages ship alongside: K-12 education, research, security and inspection, with the education and research tiers pairing the smaller units with curriculum and lab-integration software aimed at universities and STEM classrooms.
The timing is not an accident. In July the FCC added humanoid and quadruped robots to its Covered List under the Secure and Trusted Communications Networks Act, which blocks new import authorizations for those categories from Chinese manufacturers. A security contractor or critical-infrastructure operator who can no longer buy a Chinese quadruped for a US site now has a domestic option that cleared compliance certification, and Faraday Future says the Aegis was demonstrating patrol routines at IMTS in Chicago this week. The counterweight is the company's own record: a decade of missed deliveries as an EV maker, a 1-for-150 reverse stock split in July to keep its Nasdaq listing, and a robotics pivot that has yet to produce an audited deployment. The FCC decision created the opening. Whether the robots fill it is a separate question.
— Faraday Future · RobotAIGeek
🔗 Faraday Future · RobotAIGeek
What to watch next
Three of these stories are about the same physical limit. Nscale's filing puts a number on what a contract book costs before it produces revenue, Huawei's stacking design attacks the land and cooling ceiling, and Ternary Bonsai 2 shrinks the model so the building matters less. The FAA contract is the one to watch this week, because air traffic control is the first deployment where an AI prediction window of two hours carries consequences measured in lives, and the two-year pilot will produce the evidence either way. On the model side, Step 5's weight drop on October 15 is the date that matters: a 600B open MoE at one-eighth of frontier pricing resets what self-hosted inference costs.
KD Agentic publishes this digest daily. Previous editions cover model releases, agent frameworks, robotics and AI infrastructure.
Top comments (0)