Nvidia and OpenAI Eye a Half-Trillion-Dollar Ohio Data Center

1. Nvidia Is Reportedly Ready to Bankroll a Half-Trillion-Dollar OpenAI Data Center in Ohio

TL;DR: Nvidia is in talks to guarantee up to $250 billion in financing so OpenAI can lease a 10-gigawatt AI campus SoftBank is building on a former DOE uranium-enrichment site in Ohio — a project whose total cost, chips included, could top $500 billion, the largest data center deal ever floated.

Key Points:

• SoftBank's SBEnergy is developing the site ~50 miles south of Columbus; first phase (~800MW) targeted for 2028, full 10GW buildout would be roughly 3x OpenAI's existing 3.2GW Georgia campus
• Nvidia's guarantee exists because OpenAI isn't profitable and doesn't carry an investment-grade rating on its own — the GPU vendor is effectively co-signing its own customer's lease
• Japan is funding $33 billion of the power infrastructure under a "recoup investment, then cede 90% control to the US" arrangement; Commerce Secretary Howard Lutnick reportedly controls power allocation, with Anthropic, Microsoft, and Google all competing for the same capacity
• Status: early-stage talks per Bloomberg/WSJ reporting, not yet confirmed by any of the three companies — could still collapse
• Directly intersects a story below: a House committee provision that could ban Chinese-made components from data centers built on DoD land, at the exact moment companies are scouting federal sites for exactly this kind of megaproject

So What? This is the AI capex circularity story going one layer further — it's not just vendor financing customer purchases anymore, it's a chipmaker underwriting a real-estate lease so an unprofitable customer can buy the chipmaker's own product, on federal land, funded partly by a foreign government. If you're forecasting AI infrastructure demand for planning purposes, treat "in talks" project sizes as directional, not committed capacity — and watch whether Nvidia's balance sheet is quietly becoming the industry's lender of last resort.

SourcesForbes, Bloomberg

---

2. ESnet Proves Agentic AI Can Work in a NOC — While the MCP Spec Update Under It Isn't Quite Where the Headlines Say

TL;DR: The Department of Energy's ESnet published a rare, credible account of agentic AI genuinely working inside a network operations center — while the Model Context Protocol spec update its tooling depends on turns out to still be a release candidate today, not the finalized spec several outlets reported as shipping.

Key Points:

• ESnet's ORBIT (Operations Responses and Business Intelligence Toolkit) runs a centralized reasoning hub with MCP-based tool access, a semantic search layer, and outputs surfaced directly in ServiceNow — not a separate chat window nobody opens
• Tasks are built as versioned, tested "skills" rather than freeform prompting; all six initial target tasks shipped, and NOC engineers built two more on their own — genuine organic adoption, not a top-down pilot
• Multiple outlets reported MCP's biggest spec revision since launch — dropping the stateful session handshake for self-contained per-request state — as finalizing today; direct inspection of the MCP GitHub repo at research time shows only a 2026-07-28-RC tag, no final tag yet
• The changes are real once they land: MCP servers behind more than one replica can drop sticky-session routing and run behind a plain round-robin load balancer
• We flagged this exact date on July 20th, 23rd, and 27th — today was supposed to be the payoff, and as of this morning it's a payoff still in progress

So What? ORBIT is the pattern worth stealing regardless of MCP's exact ship date: skills-as-versioned-units, tool access via MCP, and outputs landing in the ticketing system your team already lives in, instead of a bespoke chat UI. Separately — if you're auditing MCP servers for the Roots/Sampling/Logging deprecation, don't start your twelve-month clock until you can see a non-RC tag in the repo yourself.

SourcesORBIT paper, arXiv, MCP release candidate blog, MCP GitHub tags

---

3. Kimi K3's Weights Are Public — and Moonshot Tightened the License, Not Loosened It

TL;DR: Moonshot AI released the full 2.8-trillion-parameter weights for Kimi K3, but the license accompanying them adds a hard commercial gate that K2's didn't have — continuing this month's pattern of "open" AI releases whose fine print doesn't match the framing.

Key Points:

• K3 is 1.56TB on Hugging Face; Moonshot itself avoids the term "open source," consistently saying "open weight" instead
• New clause: any company with more than $20 million in revenue across any trailing 12 months must negotiate a separate agreement with Moonshot before reselling K3 as a hosted inference service — K2's 2025 license only required attribution above that threshold, not a negotiated deal
• Products crossing 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3" prominently in their UI regardless of the resale clause
• Purely internal use never exposed to a third party is exempt, as is access through Moonshot's own certified partners
• Third-party providers are already listing K3 on OpenRouter at prices close to Moonshot's own — the resale gate is aimed squarely at that kind of commoditization

So What? If you're running K3 in-house, nothing changes. If you're evaluating it for a hosted product, budget for a Moonshot negotiation the moment you cross the revenue threshold — and read every "open" AI license clause-by-clause before you build a business on it. This is the third "open" claim this month (after Inkling and Block's Buzz) that didn't survive contact with its own license text.

SourcesSimon Willison, Digital Applied

---

SONiC's SRv6 Implementation Just Caught Up With the Standards Track

TL;DR: SONiC's EVPN-over-SRv6 support now covers every major EVPN route type across VPWS, VPLS, and CCC service types, while a companion IETF draft formalizes SRv6 as the single data plane for both L3VPN and L2VPN — removing a real reason to keep running MPLS-EVPN in parallel.

Key Points:

• SONiC moved from "planned capability" to running code supporting EVPN route types 1 through 5 under SRv6
• draft-ietf-srv6ops-srv6-deployment-02 (expires August 29, 2026) formalizes carrying both L3VPN and L2VPN services under one SR-native data plane
• Distinct from the SRv6 AI-fabric path-placement draft covered last week (deterministic path placement for AI backend traffic) — this is general-purpose VPN service consolidation
• SONiC's contributors are directly acknowledged in the AI-backend draft too — the open NOS is showing up as a proof point across multiple SRv6 efforts, not just one

So What? If SRv6 migration has been stuck on "our open NOS doesn't cover the route types we need," that's no longer the blocker. Worth revisiting the migration business case this quarter if it was shelved for that reason.

SourcesIETF SRv6 deployment draft, SONiC SRv6 HLD

Nvidia's Co-Packaged-Optics Switches Are Shipping — Just Not at Scale Yet

TL;DR: Nvidia's Spectrum-X photonics switches are now in limited shipment, with marketing claims like "4x fewer lasers" and "3.5x more power efficiency" all measured against unnamed "traditional" baselines — while independent supply-chain analysis says real volume doesn't arrive until 2027–2028.

Key Points:

• Top config: 512×800Gb/s or 2,048×200Gb/s ports, 400Tb/s aggregate, built on TSMC's COUPE silicon-photonics packaging
• Every headline performance multiplier compares against an unnamed "traditional pluggable-optics" baseline, not a named competing product or generation
• TrendForce reports the real constraint is optical-engine assembly yield — laser, modulator, photodetector, and coupling integration have only a handful of qualified suppliers, and CPO switches now compete with AI accelerators for TSMC's advanced-packaging capacity
• Same week, Fortinet signed with Intel Foundry to manufacture its FortiSP6 security ASIC domestically, and TSMC confirmed a fab price hike of up to 10% starting 2027

So What? Don't plan a 2026 fabric roadmap around co-packaged optics being generally available — TrendForce's independent read puts real volume 18 to 24 months out. The TSMC pricing move plus packaging capacity crunch is a good bet for why more networking silicon vendors go shopping for second-source foundries in the next year.

SourcesNvidia, TrendForce, Packet Pushers

---

Cisco's Actual Networking-Ops AI Model Is Still Not Shipped — But Now Has a Date

TL;DR: Cisco confirmed its Cloud Control platform goes live for US customers by the end of August, powered by a purpose-built "Deep Network Model" fine-tuned on routing and switching config reasoning — distinct from Antares, Cisco's already-shipped small-language-model family for vulnerability scanning that some coverage conflated with it.

Key Points:

• Cited use cases: diagnosing Wi-Fi failures by correlating security policy, AP channel config, and telemetry in one "AI Canvas"; calculating the blast radius of a planned device reboot before you execute it
• Cisco SVP DJ Sampath claims the model "solves network problems better than a frontier model," with no benchmark or eval set disclosed — model weights are slated for Hugging Face release "shortly," which will make that claim independently testable
• Pricing bundles a token allocation into existing subscriptions with metered overage still unresolved; an on-prem option exists for shops that want to avoid token costs
• Antares (Cisco's separate, already-shipped SLM family for vulnerability localization) is a different effort entirely — keep the two apart when evaluating either one
• EU/APAC rollout lags the US launch for infrastructure and regulatory reasons

So What? Nothing to test today, but mark your calendar for the Hugging Face weight drop — that's the moment "solves problems better than a frontier model" stops being a marketing line and becomes a benchmark anyone can run.

SourcesThe Register, Cisco Antares announcement

NAPALM and Netmiko Both Shipped Real Releases This Week

TL;DR: NAPALM cut its first release in nearly a year (5.1.1 then 5.2.0) fixing long-standing BGP and NX-OS bugs, while Netmiko 4.7.0 added twenty new device drivers spanning optical transport gear and IPInfusion OcNOS.

Key Points:

• NAPALM 5.2.0 fixes a BGP-origin-code crash in get_route_to() and corrects is_up/is_enabled field mapping on NX-OS — both real bugs shops may have been quietly working around
• NAPALM also modernized its toolchain: ruff replacing pylama, uv replacing requirements.txt, GitHub Actions hash-pinning for supply-chain hardening
• Netmiko 4.7.0 added drivers for Fujitsu, Smartoptics, Perle, Nokia, Aviat, APC, OpenGear, and IPInfusion OcNOS, plus an IOS-XR commit replace option
• Netmiko pinned Paramiko below version 5.0 after upstream breaking changes — worth checking your lockfiles before a transitive upgrade breaks your automation

So What? If you've been maintaining a workaround for either NAPALM bug, or scripting raw Paramiko against optical transport or OcNOS gear, both fixes are worth the version bump this week.

SourcesNAPALM releases, Netmiko 4.7.0

---

NVIDIA Shrinks Its Quantum-Calibration Model to Single-GPU Size — On Its Own Benchmark

TL;DR: NVIDIA's Ising Calibration 1.5, an open vision-language model that interprets quantum processor diagnostics, is now 31 billion parameters with a first-ever NVFP4-quantized version that fits on a single consumer GPU — with all the "beats closed models" claims measured against NVIDIA's own QCalEval benchmark.

Key Points:

• Claims an 86.68% improvement over its predecessor on QCalEval with in-context examples, roughly 10% better zero-shot than comparable open models — no third-party baseline for either number
• The NVFP4 quantization recipe (selective BF16 layers, Random Hadamard Transforms, stochastic-rounded gradients) is the genuinely reusable part, independent of the quantum-calibration use case
• Licensed under the Linux Foundation's OpenMDW License — same permissive lineage as Nemotron 3 Ultra, and one of the few "open" claims this year that actually holds up
• Claimed competitive with Gemini 3.5, Claude Fable 5, and GPT 5.6 Sol, but no numeric head-to-head is published

So What? The pretraining and quantization technique is worth studying regardless of whether the quantum-calibration benchmark holds up — treat the "beats closed models" claim as unverified until someone runs an independent comparison.

SourcesNVIDIA Technical Blog, NVIDIA Build

---

Texas Just Built the First Real "Data Centers Pay Their Own Way" Mechanism

TL;DR: Following Governor Abbott's directive, ERCOT's new "Batch Zero" process evaluates large-load data center proposals of 75 megawatts or more collectively rather than project-by-project, allocating transmission capacity and identifying required grid upgrades through 2032 — with developers required to post financial security and start paying transmission charges before they even energize.

Key Points:

• PUCT has already adopted standardized large-load forecasting rules and blocked existing generation from being redirected to new data centers
• Batch Zero determines available capacity by location; future batch studies are expected to become a permanent annual or biennial process starting 2027
• A bigger rulemaking finalizing around December 2026 would require financial security before developers even enter ERCOT planning studies
• PJM and MISO are reportedly watching Texas as a template for their own large-load queue problems

So What? This is the first grid operator to move past "data centers should pay their share" rhetoric into an actual collective-capacity-allocation mechanism. If you're siting anything in ERCOT territory, Batch Zero eligibility just became a real go/no-go planning input, not a future concern.

SourcesData Center Knowledge, Texas Governor's Office

A Sixth Siting-Friction Axis: Supply-Chain Compliance on Federal Land

TL;DR: A House Armed Services Committee provision would bar the Pentagon from leasing land to data center operators unless they exclude Chinese-, Russian-, Iranian-, or North Korean-made components — and the Army itself is objecting, warning the domestic industrial base can't yet supply everything the rule would require.

Key Points:

• Passed committee only, in the House version of the FY2027 NDAA — still needs full House passage and Senate reconciliation
• The Army calls it a "federal land penalty," since it applies stricter sourcing standards to military bases than identical projects built on private land
• Lands the same week as the Nvidia/OpenAI/SoftBank Ohio project on a former DOE site — federal land is actively being courted for exactly this kind of megaproject

So What? This joins power, land-opposition, labor, water, and transmission as a genuine sixth siting-friction axis — supply-chain and national-security compliance. If a project is scouting federal land, sourcing compliance now belongs on the same due-diligence checklist as power interconnect queue position.

SourcesDataCenter Dynamics, Federal News Network

---

Physicists Build a Complete Quantum Computing Toolkit Out of Particles That Remember Being Swapped

TL;DR: A Quantinuum-led team assembled the first fully universal quantum gate set from topological operations alone — braiding and fusing non-Abelian anyons, quasiparticles whose quantum state depends on the path they took around each other, not just where they started and ended.

Key Points:

• The team entangled 54 trapped-ion qubits on Quantinuum's H2 processor into a topological state hosting anyons in an S3 model
• Braiding alone was proven insufficient for universal computation in the early 2000s; the team implemented a 2003 proposal combining braiding with fusion-based measurement to close that gap, sidestepping the resource-heavy "magic state distillation" conventional fault-tolerant approaches rely on
• Published in Nature, peer-reviewed
• This is the topological approach Microsoft has bet its entire quantum program on for over a decade, largely without a hardware payoff — a trapped-ion company demonstrating it validates the underlying concept independent of whose hardware you're betting on

So What? Not deployable today, but a real validation of a second, distinct route to fault-tolerant quantum computing beyond the error-correction-heavy approaches most of the industry is chasing. Worth a bookmark if you track quantum computing roadmaps for infrastructure planning.

SourcesNature, The Quantum Insider

Engineers Watch Heat Travel in Straight Lines at Room Temperature

TL;DR: UCLA researchers observed "phonon focusing" — heat traveling in directional, ray-like beams rather than diffusing uniformly — in boron arsenide at room temperature, a wave-like heat-transport behavior previously only seen at cryogenic temperatures.

Key Points:

• Different crystal facets produced distinct six-fold, eight-fold, and four-fold focusing patterns depending on lattice orientation, with coherent directional transport persisting over roughly a micrometer
• Published in Nature Physics, peer-reviewed
• Researcher framing: "enabling heat to be guided, focused and redistributed with nanoscale precision at room temperature" lays groundwork for what they call quantum thermal engineering

So What? The most infrastructure-adjacent science story this cycle. If heat can eventually be steered like light through fiber instead of just diffusing, that's a long-term lever on the thermal bottlenecks driving liquid cooling and rack redesign in AI datacenters — worth filing next to today's Texas grid and Nvidia/OpenAI power stories as a reminder the thermal side of the AI buildout has its own frontier research pipeline.

SourcesNature Physics, UCLA CNSI

---

• NVIDIA's Cosmos-H-Dreams trains surgical robots against a real-time generative simulator instead of real tissue — a teacher-student distillation pipeline gets it to 160 frames per second on a single RTX PRO 6000, learning from failed suturing attempts as well as successful ones.
• A new arXiv study finds prompt wording measurably changes on-device LLM energy draw — imperative-verb choice and instruction phrasing shift decoding length, and therefore real, measured battery consumption.
• Compass Datacenters priced $413 million in asset-backed bonds across three tranches backed by seven operational facilities; only the smallest tranche, at $50 million, actually got an S&P rating — an A, on a 1.69x debt service coverage ratio.
• ServeTheHome's independent review of an 8-GPU NVIDIA HGX B300 server clocked 30 kilowatts of power capacity in a single 4U chassis with a choice of two liquid-cooling paths — a concrete look at what's driving 200-plus-kilowatt rack density conversations.
• Nebius closed on 600 acres in Schuylkill County, Pennsylvania for a 1.2-gigawatt campus, designing in closed-loop liquid and air cooling specifically to reduce water draw — a facility proactively built around this month's water-siting-friction thread rather than fighting it after the fact.
• A new arXiv paper proposes a Huffman-code-based addressing scheme for quantum networks that preserves quantum coherence while routing — demoed on a thirteen-node test network with claimed perfect fidelity.

SourcesHugging Face, arXiv (on-device LLM energy), DataCenter Dynamics, ServeTheHome, DataCenter Dynamics (Nebius), arXiv (quantum addressing)

---

• Cisco's Hugging Face weight drop for its networking-ops "Deep Network Model" — the moment its "beats a frontier model" claim becomes independently testable
• Whether the MCP spec actually gets a final (non-RC) tag today, and what changes between the RC and the final text
• The Nvidia/OpenAI/SoftBank Ohio deal — still "in talks," worth watching whether it firms up or quietly collapses
• The House NDAA provision on Chinese components at DoD-land data centers, as it heads toward Senate reconciliation

---

• Domains researched: 6 (network architecture, network automation, AI/ML, datacenter, security, science)
• Web searches: ~20 across all agents, supplementing a moderate RSS digest (81 articles, 22 feeds)
• Items published: 12 major items + 6 quick takes
• Security: no significant architecture updates this cycle (checked and confirmed dry, not skipped)
• Quality score average: 4.5/5