# China's 1,000 km Hollow-Core Fiber Trial: What It Means for AI Data Center Networks > Source: > Published: 2026-11-08 00:00:00+00:00 **Short answer:** On **November 4, 2026**, [China Telecom Research Institute](https://www.chinatelecom.com.cn/ct/news/jtxw/169343.html?ref=explainx) said it completed what it calls the **world''s first 1,000-kilometer bidirectional real-time transmission test** on **hollow-core** (empty-core) fiber, with **YOFC** and **ZTE**. Over **1,007.48 km**, the trial reported **800 Gb/s per wavelength** in **both directions** on a **single fiber**, **128 Tb/s** aggregate in the extended **C+L** band, and officials tie the result to **lower-latency backbones for China''s compute networks**. State broadcaster **CCTV** amplified the story on **November 7–8**, which is why the trial resurfaced in AI infrastructure feeds days after the press release. This is not a model launch. It is **plumbing** — the kind of milestone that only matters if you run **multi-site GPU clusters**, **disaggregated inference**, or **synchronous training** across regions. For that audience, shaving **milliseconds per kilometer** and doubling **bits per strand** can change how many parallel fibers you must trench between campuses. ## What the official release claims The [China Telecom news item](https://www.chinatelecom.com.cn/ct/news/jtxw/169343.html?ref=explainx) lists three headline results: 1. **Longest claimed point-to-point hollow-core real-time span** at this scale, with**at least 3 dB** loss margin per hop and**at least 3 dB OSNR** maintenance margin called out. 2. **First 1,000 km-class bidirectional transmission on one fiber** , with reflection-related OSNR penalties between**0.5 dB and 4 dB** depending on conditions. 3. **Extended C+L band** without the**CO₂/CO gas line absorption (GLA)** problem that had been degrading hollow-core performance — addressed through fiber process tweaks, conventional amplifier choices, and pre-equalization on the transmit side. Engineering readers should notice the vendor stack: **fiber from YOFC**, **systems from ZTE**, **operator lab from China Telecom**, on national **key laboratory** platforms. That is the usual Chinese pattern for declaring a **standards-relevant** win before commercial volume. ## How the team handled the engineering hurdles The release describes three design choices that explain why this span was hard. First, the **gas line absorption** problem: **CO₂ and CO absorption lines** inside the extended C+L band degraded earlier hollow-core links, and China Telecom calls removing them the core technical challenge. The team tuned **fiber fabrication** so the band stays free of those lines, and it describes the result as the longest GLA-free extended C+L hollow-core link reported anywhere. Second, the team did **not** rely on exotic amplifiers. Because of hollow-core loss and amplifier behavior, it avoided high-power units and used **commercial amplifiers at conventional output power**. That matters for operators, since it suggests existing line equipment could carry such links. Third, a **lightweight transmitter-side pre-equalization** scheme kept channel performance balanced across the whole band. The team also built an **end-to-end model** of reflections and inter-modal interference, then checked that losing a band during maintenance caused **no measurable degradation**. ## Hollow-core versus solid-core in plain language Today''s internet runs overwhelmingly on **solid-core** fiber: light moves through **silica glass**, typically near **two thirds** the speed of light in vacuum because of the refractive index. **Hollow-core** fiber routes light largely through **air or gas** in the center. [CCTV''s summary](https://www.chinanews.com.cn/gn/2026/11-07/10728184.shtml?ref=explainx), citing **Wang Ruichun**, YOFC''s chief scientist, contrasts **about 200,000 km/s** effective travel in glass with speeds **closer to 300,000 km/s** in air. Executives quoted in the same coverage describe **about 30 percent lower latency** versus traditional solid-core links **when the design fits**. Capacity claims are equally loud: state media says the trial **doubles transmission capacity** versus the best-performing solid-core comparisons they used, and **extends distance by about 60 percent** relative to that baseline. Independent third-party benchmarks are not in the press release; treat those ratios as **vendor trial comparisons**, not a global leaderboard. ## Why AI infrastructure watchers care Large training runs already treat **the network between racks and buildings** as a bottleneck. Techniques like **disaggregated inference** — pushing **KV cache** or activations across machines — punish **slow or lossy** links. Our write-up on [vLLM push-based KV transfer](https://explainx.ai/blog/vllm-push-based-bidirectional-kv-transfer-disaggregated-inference-aws-p5en-2026) shows how hyperscalers measure **milliseconds** of added tail latency when pools are not co-located. China''s domestic build-out makes the timing legible. explainx.ai has tracked **multi-gigawatt campuses** such as [Envision Galaxy in Inner Mongolia](https://explainx.ai/blog/envision-galaxy-campus-wind-powered-ai-data-center-ulanqab-inner-mongolia-2026) and national **GW-scale shell capacity** comparisons in [SemiAnalysis reporting](https://explainx.ai/blog/china-24-gw-compute-vs-us-56-gw-semianalysis-inner-mongolia-2026). Fiber between those sites is the **long pole** once land and power are booked. **Huo Xiaoli**, a China Telecom expert quoted by CCTV via [Global Times](https://www.globaltimes.cn/page/202611/1373868.shtml?ref=explainx), frames the trial as proof hollow-core can **reuse existing transmission gear** while delivering **long-distance, high-capacity, low-latency** links — reserving technology for **算力网** (compute-power network) construction. **Wang Ruichun** adds that **training and inference** need **tight coordination across data centers**, so **bandwidth and delay** between facilities directly affect **GPU efficiency** and **hardware spend**. That is the AI angle: not that fiber trains models, but that **slower east-west traffic** forces **more GPUs sitting idle** waiting on peers. A green stopwatch beside pipeline elements suggests latency-sensitive links between facilities *Image: explainx.ai library — millisecond-scale backbone delay matters for distributed training and inference pools.* ## Bidirectional 128 Tb/s on one fiber Headlines mix **per-direction** and **aggregate** numbers. China Telecom''s release emphasizes **800 Gb/s PCS-16QAM** **per wavelength** **in each direction** on **one physical fiber**, totaling **128 Tb/s** in the extended C+L band. [CCTV reporting](https://www.chinanews.com.cn/gn/2026/11-07/10728184.shtml?ref=explainx) translates that into a consumer metaphor: **about 64 Tb/s each way**, roughly **160 4K movies per second** at **50 GB** each, and a **one-way 1,000 km** hop in **about 3.43 ms**. Bidirectional DWDM on a single hollow-core strand is technically fussy because **back-reflection** and **modal interference** can chew OSNR. China Telecom says it modeled **end-to-end reflection** and **mode coupling**, ran **band-drop** fault scenarios, and still saw **no performance collapse** — the kind of evidence operators want before they trust one fiber for both directions on a backbone route. ## Deployment reality: trial versus trenches Laboratory records do not automatically become **metro fiber counts**. The same CCTV package notes a **commercial hollow-core financial dedicated line** between **Shenzhen and Hong Kong**, plus **more than ten** commercial or pilot projects in **Asia and Europe** mentioned at a high level. What is missing from primary documents we checked: a **public per-kilometer cost**, a **national rollout calendar**, and **independent replication** outside the China Telecom–YOFC–ZTE consortium. [Frontword''s summary](https://frontword.com/innovation/hollow-core-optical-fiber-800g-test/?ref=explainx) repeats the technical figures and likewise notes **no disclosed commercial timeline**. For builders outside China, the near-term takeaway is **directional**: hollow-core is moving from **short spans and specialty finance links** toward **kilometer-class backbone claims**. Whether your cloud region benefits depends on whether your provider buys those cables on routes you use. ## How this sits next to other fat-pipe projects Subsea projects still grab the **transcontinental** headlines. Meta''s [Petal cable](https://explainx.ai/blog/meta-petal-subsea-cable-petabit-2026) targeted **petabit-per-second** Atlantic capacity. The China Telecom trial is **terrestrial**, **bidirectional on one strand**, and anchored in **AI policy language** about **compute networks** rather than consumer streaming. Inside the data hall, different bottlenecks dominate — **NIC speeds**, **NCCL collectives**, **rack power**. But when clusters span **tens to hundreds of kilometers**, **backbone fiber** returns to the critical path. That is the same reason [PJM wholesale power analytics](https://explainx.ai/blog/pjm-data-centers-11-dollars-mwh-wholesale-capacity-monitoring-analytics-2026) and **siting fights** now share headlines with **model weights**: scaling is **systems all the way down**. A green cable links a transformer to a small server cabinet, tying power and network constraints for large compute sites *Image: explainx.ai library — AI sites face coupled power and connectivity planning.* ## What to watch next Credible follow-ons would include **third-party field trials** on non-consortium fiber, **priced offerings** from **YOFC** or **ZTE** with **km-scale SLAs**, and **cloud provider statements** that hollow-core is on **specific AI campus routes**. Without those, treat November''s wave as **a verified milestone**, not a procurement mandate. If you operate **multi-region inference**, none of this replaces **application-level** latency work — caching, batching, and [reranker placement in RAG](https://explainx.ai/blog/how-to-add-reranker-rag-pipeline-latency-models-guide-2026) still dominate user-visible delay. Backbone wins matter when you are already **network-bound between your own halls**. ## How to read vendor trial numbers as a builder Optical press releases stack **capacity**, **distance**, and **modulation format** in one breath. A disciplined reader separates them. **Distance** answers whether the fiber survives **amplifier spacing** and **OSNR** on a real route, not just in a loopback box. China Telecom''s release emphasizes **maintenance margins** and **fault scenarios** for that reason. **Capacity** answers how many **wavelengths** fit at **800G** class rates. DWDM economics mean operators care about **spectral efficiency** per fiber because trenching new conduit is slow. **Latency** answers whether **synchronous training steps** or **tight inference SLAs** can span more kilometers without extra copies of weights. A **30 percent** latency improvement on paper does not automatically flow to your GPU job unless your bottleneck is literally **speed-of-light limited** on that hop. If you are not procuring backbone fiber yourself, the actionable output is simpler: ask your **cloud or colocation provider** whether long-haul routes use **hollow-core pilots**, what **RTT** they measure between regions you use, and whether **cross-AZ traffic** is priced differently on those paths. That question costs nothing and surfaces whether a trial like this one affects your bill. Record the answers when you plan **multi-site fine-tunes** or **disaggregated inference** layouts so network assumptions stay attached to architecture docs, not hallway folklore. Revisit the notes when backbone vendors publish the next distance record. ## Related reading **Primary sources:** [China Telecom, November 4, 2026 release](https://www.chinatelecom.com.cn/ct/news/jtxw/169343.html?ref=explainx) · [Science and Technology Daily republication](https://www.stdaily.com/web/gdxw/2026-11/04/content_607916.html?ref=explainx) · [CCTV / China News summary, November 7, 2026](https://www.chinanews.com.cn/gn/2026/11-07/10728184.shtml?ref=explainx) *Published November 8, 2026. Trade press cited for executive quotes where English translations exist; we did not attend the trial.*