CoreWeave shares fell 9.83% Thursday at the close on Nov. 5 as investors worried about a backlash against data centers after votes in several states, CNBC reported, nine days after the company announced a global AI backbone network without publishing speed, latency or capacity measurements for it. The Oct. 27 announcement covered most of a 22.5MW deployment now online in Copenhagen and seven new network points of presence (PoPs) across the Middle East and Asia Pacific.

For AI teams choosing where to train and serve models, the question is which parts of that announcement are verified specifications and which are company claims. This article separates the two, and explains what link-level encryption and InfiniBand networking do and do not guarantee.

Texas, Ohio and Georgia Democrats Promised Data Center Pauses

Fortune reported on Nov. 4 that Democrat Gina Hinojosa defeated Texas Gov. Greg Abbott after promising during the campaign to impose a data center construction moratorium and eliminate subsidies and tax breaks for major developers. In Ohio, Amy Acton called for a conditional moratorium unless developers cover all utility costs, use union labor and meet environmental standards, and in Georgia, Keisha Lance Bottoms promised a temporary statewide pause on new data center construction. Fortune also reported that shares of CoreWeave, Applied Digital and Cipher Digital fell on Nov. 4.

Hinojosa’s July proposal, as posted on her campaign website, called for a moratorium on new data centers until the Legislature convenes and a ban on nondisclosure agreements that hide water and energy use. That source is the campaign’s own account of its position. Fortune reported that Abbott had directed state regulators to develop measures making data centers cover the cost of the energy infrastructure they use, and later instructed regulators to pause new data center permits until audits of electricity and water use were completed. Fortune also noted that governors hold more direct tools than Congress, including permitting, regulation and tax incentives, though some actions may need legislative approval.

CoreWeave operates a $1.6 billion data center in Plano, Texas, built for Nvidia. The market reaction reflects uncertainty about future approvals, not an enacted ban.

Copenhagen Site Brings Most of 22.5MW Online With Nvidia Blackwell Systems

CoreWeave said it brought most of a planned 22.5MW of IT capacity online at atNorth’s DEN01 data center in Copenhagen, its first deployment in Denmark. CoreWeave has not said how much remains to come online. The site runs Nvidia Blackwell systems connected with Nvidia Quantum-X800 InfiniBand, and it lifts CoreWeave’s European footprint to nine data centers across five countries.

DEN01 has a heat reuse agreement that sends surplus heat into the local district heating network. atNorth CEO Eyjólfur Magnús Kristinsson was quoted in the same release on the facility.

Seven Network Points of Presence Begin Later in Q4 2026, Full Loop by Mid-2027

The new PoPs are planned for the United Arab Emirates, India, Indonesia, Singapore, Australia, South Korea and Japan, connected over terrestrial and subsea fiber. They begin coming online later in the fourth quarter of 2026, with a full loop expected in mid-2027. CoreWeave’s blog says the backbone today connects 51 active data centers, including eight added in 2026, across North America and Europe.

The India PoP will connect 240MW of capacity in Navi Mumbai, and the Indonesia PoP will tie in the three facilities CoreWeave announced in August. CoreWeave reported more than 1.5GW of active power as of June 30, 2026, and about 4.2GW of contracted power, which is capacity secured but not yet running, in figures its blog attributes to the Aug. 11 earnings call.

MACsec Encrypts Link by Link, So Every Hop Must Be Trusted

CoreWeave says it runs MACsec (IEEE 802.1AE) encryption in hardware at line rate on three surfaces: backbone links between regions, Direct Connect traffic arriving from customer networks, and traffic inside its data centers. The company says MACsec is not always supported end to end at other providers, but it does not name them.

MACsec operates at the media access control layer, Layer 2, which means it protects Ethernet frames on each link rather than across a routed path. One technical explainer says MACsec operates on a hop-by-hop basis, so each switch or endpoint on the path must be trusted and correctly configured, and security depends heavily on key management. The same source says the encryption runs in network card and switch hardware at line rate with negligible added latency, and that MACsec frames grow by roughly 24 to 32 bytes.

The practical consequence for customers is that CoreWeave’s claim of coverage on every link is a claim about configuration across its whole fleet. It is not a mathematical property of MACsec. Buyers handling valuable model weights can ask who holds the keys and how each device in the path is secured.

Elephant Flows Explain Why CoreWeave Built Multiple 400Gbps Links per PoP

Each PoP carries multiple 400Gbps connections over a unified optical IP network, and CoreWeave says 10Gbps and 100Gbps links remain common in cloud connectivity, without offering data for that comparison. AI traffic stresses networks because of the elephant flow: after a GPU cluster finishes a batch, it sends all of its output at once, driving network utilization toward 100% during the transfer.

Because AI training is synchronous, one delayed transfer holds up the jobs that depend on it. That is why AI fabrics inside data centers are typically built to match server-facing and spine-facing capacity, while non-AI networks may run 2:1, 3:1 or 4:1 oversubscription. Resilience is the third design goal. CoreWeave says each PoP links to at least two other PoPs over four physically diverse paths, and each region is dual-homed, though its blog does not make clear whether the four paths apply per PoP or network-wide.

Does Quantum-X800 InfiniBand Make Copenhagen Different From Rivals?

Nvidia’s March 2024 release called Quantum-X800 and its Ethernet sibling the world’s first platforms capable of end-to-end 800Gb/s throughput, with 5x the bandwidth of the previous generation and a 9x increase to 14.4 teraflops of in-network computing. Dell’s datasheet for the Q3400-RA model lists 115.2Tb/s of throughput across 144 non-blocking 800Gb/s ports. CoreWeave names only “Quantum-X800” for Copenhagen, so which switch model the site uses is not confirmed. Neither Nvidia nor Dell publishes a numeric latency figure in the documents reviewed.

InfiniBand connects the Blackwell systems at the Copenhagen site, while CoreWeave describes its backbone between regions as a unified optical IP network. InfiniBand has a single dominant supplier: Nvidia acquired Mellanox, described as the last independent InfiniBand supplier, in 2019. Fierce Network reported that analyst Alan Weckel of 650 Group called it too early to project market share, while a Cisco Fellow said the industry is moving quickly toward Ethernet.

Microsoft Described Dedicated AI Wide-Area Network in November 2025

CoreWeave calls its backbone the first global network purpose-built for AI workloads, a claim its blog does not support with any comparison to competitors. On Nov. 12, 2025, Microsoft’s Scott Guthrie described a dedicated AI WAN optical network extending Fairwater’s networks, and said Microsoft delivered over 120,000 new fiber miles across the United States last year. Microsoft’s design uses a two-tier, Ethernet-based backend network and its own SONiC switch operating system.

Whether CoreWeave’s “first” holds depends on how “global” and “purpose-built” are defined. CoreWeave’s network connects North America and Europe today, and the other four continents named in its announcement are still planned.

Petabit-Scale, 96% Goodput and 20% Higher MFU Lack Public Baselines

CoreWeave describes the finished network as Petabit-scale and equates that to streaming a 4K movie to 66 million people at once, without stating a bitrate or assumptions. It has not published aggregate capacity, interregional latency or measured transfer rates.

The company cites 20% higher Model FLOPs Utilization (MFU) and 96% goodput from its whitepaper “Purpose-Built Clouds for AI at Scale,” but its blog post does not state what those figures are measured against. In networking, goodput normally means useful application-level data delivered per unit of time, excluding protocol overhead and retransmitted packets. CoreWeave applies the term to productive GPU time, so readers should check which definition the whitepaper uses. The blog also claims a two-tier fabric connecting up to 128,000 GPUs and up to 8x lower read latency from its LOTA storage technique, again without a stated baseline.

Debt of About $35 Billion Raises Stakes for Network Buildout

The Motley Fool reported on Nov. 3 that CoreWeave carries approximately $35 billion in debt and has built a $104 billion revenue backlog. Bisnow reported in October that obligations total close to $50 billion when $16.3 billion in data center lease commitments are included, and that analysts disagree: Andromeda Capital’s Alberto Gallo said CoreWeave would need a capital injection or a bailout, while Goldman Sachs analyst Gabriela Borges upgraded the stock, arguing AI demand will justify the expansion costs. Bisnow also reported that CEO Michael Intrator wrote in an internal email that the market misunderstands and acts irrationally, and described the company’s model as more than supporting its debt.

Chief operating officer Sachin Jain, who previously held senior roles at Oracle Cloud, Google Cloud and Amazon according to CoreWeave’s earlier hiring announcement, said in the Oct. 27 release that customers need more than access to computing capacity and should be able to focus on building AI rather than managing network complexity.

CoreWeave’s backbone is a bet that dense, encrypted, redundant long-haul capacity will matter as much to AI customers as GPU supply. The verified parts are the Copenhagen capacity, the seven planned PoPs, the multiple 400Gbps links per PoP and the MACsec and redundancy design, all as stated by CoreWeave. The unverified parts are the Petabit-scale figure, the interregional performance and the MFU and goodput gains. Added to that is a financing and permitting environment that shifted within a week. The first new PoPs are due later this quarter, and measured transfer rates would be the clearest test.

Frequently Asked Questions

Does MACsec encrypt data from end to end?

MACsec secures each link between connected devices, and traffic is decrypted and re-encrypted at every hop, so protection depends on every device in the path being trusted and keyed correctly. One explainer states that MACsec does not provide end-to-end confidentiality across a multi-hop network.

Has anyone independently verified CoreWeave’s network performance claims?

CoreWeave has not published aggregate capacity, interregional latency or measured transfer rates, according to StorageReview’s report on CoreWeave, and its MFU, goodput, GPU-count and storage-latency figures come from its own materials with no stated baseline.

What is a backbone network for AI?

A backbone network interconnects other networks and usually has greater capacity than the networks it links. CoreWeave’s version links its data centers over multiple 400Gbps connections per PoP to move large, long-running AI data transfers between regions.

When will CoreWeave’s global AI network be complete?

The seven new PoPs begin coming online later in the fourth quarter of 2026, and CoreWeave expects a full global loop in mid-2027, according to its Oct. 27 press release.