MicroLED 苦等十年的量产线,可能先被一批“不按寸算钱”的客户订走After a decade-long wait, MicroLED's mass production line may be snapped up first by clients who aren't charged by the inch.


AI free token/AI免费积分
 

给美军做瞄准镜的公司,开始给 GPU 供货了

先说一家你可能没听说过的公司:Kopin。

它做了几十年的军用微表明器——头盔表明器、热成像瞄准镜、单兵穿戴设备。一句话概括:给士兵看的显示屏。

而它现在的主推机型叫Neural I/o,客户不是士兵,是数据中心。

时间线是这样的:

2026 年 4 月,Fabric.AI 向 Kopin 下达1500 万美元初始开发订单,资助 Neural I/o  演示芯片组。作为对价,Kopin 拿到 Fabric.AI19.9% 股权,并成为该芯片组的独家制造商。

9 月 17 日,双方聘请律所,推进联合知识产权布局。

工艺原理说破了很朴素:把可编程的 MicroLED 像素,重新外观成超高速光收发器,用光子而不是电子搬数据——替换掉 AI 数据中心里  GPU 之间的铜线。

MicroLED 苦等十年的量产线,可能先被一批“不按寸算钱”的客户订走

双方联合白皮书给出的测算模型表明:在一个10 万个 GPU的集群里,MicroLED  光互连相比传统光学思路可减少约75%的相关功耗,净节省约30MW持续电力需要。

30MW 意味着什么?不只是每年数千万美元的电费和空调费。更现实的意义是——运营商可能因而不用新建变电站,也不用因为电网并网受限,把 AI  项目往后推。

给夜视仪做显示屏的工艺,跑去应对 GPU 集群的功耗墙,听起来有点荒诞。

但它恰恰暴露了 MicroLED 这十几年最尴尬的处境。

苦等十年:它卡在"一颗都不能错"

MicroLED 被叫做"下一代表明工艺"十几年了。亮度高、功耗低、寿命长、不怕烧屏,纸面参数几乎全面碾压 LCD 和 OLED。三星、LG  年年在展台上端出新东西,从电视到手表到车载,年年惊艳。

但十几年过去,它主要还是活在展台上。

卡在量产。

MicroLED 芯片没法直接在表明背板上长出来,得先在蓝宝石、砷化镓这类衬底上用 700°C  以上的工艺制备,然后一颗一颗搬到背板上。

搬多少?一台 8K 电视,7680×4320 像素再乘以三色,大约1 亿颗微米级芯片。

换个说法感受一下:如果每秒装 1 颗、24 小时不停,要装 3 年多。

而这还只是数量难题。更要命的是第二条——

一颗坏点,就毁掉整屏。

这就是为什么它苦等了十年:不是做不出来,是做出来不够便宜、也不够稳定。工艺跑不顺,良率上不去,成本下不来。

于是难题来了:同一颗芯片,凭什么换到数据中心就容易了?

三个在表明里要命的缺点,在通信里都不是难题

答案是:评价标准整个翻了个面。

MicroLED 在表明上有三个致命痛点,而这三个,在通信情形里几乎都不是难题。

第一,颜色。表明必须 RGB  三色,三色的亮度和波长还得匹配,光是分色对位就够喝一壶。通信呢?单色就行——主流思路是同色通道并行,直接跳过最麻烦的环节。

第二,坏点。表明上一颗坏点肉眼可见,不可接受。通信可以外观冗余通道,坏一个换一个,使用者毫无感知。

第三,亮度均匀性。表明上的 mura(亮度不均)不可接受,为了挑出一致性够好的芯片,分 bin  的成本极高。通信关心的是误码率和功耗,对亮度均匀性并不敏感。

而下面这组表现参数,在通信情形里每一项都正中靶心:

每比特能耗 1–2 pJ/bit,在同等速率与距离下约为铜缆思路的5%;

耐 125°C 高温,寿命超3 万小时。因为基于自发辐射,天生不怕热,所以不需要 TEC 温控;

因为是并行低速架构,还能省掉 DSP、CDR 这些高功耗器件。平台复杂度和成本,一起往下掉。

业界管这种思路叫"宽而慢"——不追求把单通道做得越来越快,而是用几百上千个低速通道并行,靠数量换总带宽。

微软的 MOSAIC 原型,据报道已在 20 米距离上完成单通道 2Gbps 稳定传输;Avicena 在 OFC 2026  现场演示了512Gbps链路,并计划推进到 896Gbps。

于是就有了全文最核心的那句话:

MicroLED 在表明端最要命的"全彩难、坏点致命、均匀性苛刻",到了通信端几乎都不是难题。

它在表明里被骂了十年的短板,是它在 AI 时代的入场券。

当然得说清楚:这不等于通信端就"容易"。1mm² 里集成数百个通道,换算下来节距也在几十微米量级,密度要求仍旧不低。

真正的松绑来自那四点——单色、可冗余、不必做分色对位、坏点不影响观感。叠在一起,工程难度是断崖式下降的。

面板卖的是"寸",光互连卖的是"瓦"

工艺上说得通了,还有一道更现实的关:谁付钱?

这就得说到表明行业的一个基本游戏规则。

面板是怎么卖的?按尺寸、按片计价。报价单上写得明明白白:32 英寸多少、55 英寸多少、65  英寸多少。客户拿着单子,比的永远是"每一寸多少钱"。

所以面板厂唯一的生存法则就是——让单位成本随尺寸和世代线不断下降。做不到,就出局。

好,现在换到数据中心这边。

云厂商买光互连,问的第一个难题不是"这一片多少钱"。他们算的是另一笔账:每搬 1  比特数据要花多少电?整机散热扛不扛得住?为了多出来的这点功率,要不要新建一座变电站?

这就是标题里那句"不按寸算钱"。

这个差别,带来一个格外关键的推论:

一颗 MicroLED  芯片,在面板客户眼里是成本项——是要被拼命压价的对象;到了数据中心客户眼里,它变成了省电工具——是能帮他省下电费和变电站投资的东西。

一个是被压价的对象,一个是能参与分成的思路。

这就是为什么"不按寸算钱"的客户,有可能在 MicroLED 成本还没打平的时候,就先下订单。

注意是"有可能"。当下多数项目还在送样验证,产业界普遍判断 2027 年前后才到成本拐点。这个窗口能不能真正吃到,后面会细说。

把芯片"冲"进坑里:工艺这边,最近刚好有一场会

再来看工艺。

就在 9 月 17 日,MicroLED Connect 2026 在荷兰埃因霍温召开。议题恰好回答了"良率和成本到底卡在哪"。

路线一:把"转移"和"键合"压缩成一步

韩国电子通信研究院(ETRI)展示的工艺叫SITRAB(同步转移键合)。

传统工艺是"先转移、后键合"两步走。SITRAB 用一种自研的无溶剂胶,涂得极薄且均匀,在 980nm  红外激光照射数秒后,一步同时完成转移与焊接。

真正巧妙的是"一次涂覆、多次加工":这种胶经过 6 次 300W、6 秒的激光辐照后,核心官能团仍保持活性——要到 120°C  以上才完全固化。听起来像个材料学细节,但它换来三个表明端梦寐以求的能力:

可拼接。15×15 阵列分四次拼成 30×30,无缝扩展,直指大尺寸面板。

可修复。把冗余芯片转移到修复电极就行,不用移除坏芯片。研究团队公布的数据是,冗余修复能把故障面板良率从 83.01%  提高到 99.80%。

可异质集成。不同材质、颜色、厚度的芯片(例如红 9μm、绿/蓝 6μm)能依次集成到同一块背板上。

论文展示的成果涵盖 165 PPI 的 AlGaInP/InGaN 混合背板,以及 32×32 全彩面板(324μm 像素节距、50μm  子像素节距)。

路线二:用狭缝涂布把芯片"冲"进陷阱

英国谢菲尔德大学的衍生公司Pixel-Flo走的是另一条路。

它的判断很直接:机械式 pick-and-place  有根本上限——方形载具和圆形晶圆不匹配,边缘的芯片白白浪费;更要命的是,单位成本不随基板尺寸增大而下降。在面板行业,这一条几乎是判死刑。

Continuous-Flo  的做法很有画面感:把已知良品芯片释放进"墨水",用标准光刻在基板上做出一个个陷阱,再用改良过的狭缝涂布(slot-die)工艺,靠弯月面力把芯片连续引导进陷阱。已经演示了4μm  芯片、25μm 间距的沉积。

好处很直白:边缘芯片也能用、能用更小也更便宜的芯片、工序更少、吞吐更高、单位成本随基板尺寸下降、还能单步完成 RGB。

公司 2025 年从谢菲尔德大学分拆,2026 年 7 月完成525 万英镑(约 4780  万元人民币)种子轮,还专门请了驻中国台湾的业务发展负责人对接亚洲面板供应链。

会上同台展示的还有 TU Ilmenau  等其他思路。几条路线并跑,说明工程路线还没收敛——但也说明业界已经不在讨论"能不能做",而是在争"哪条更便宜"。

一扇被忽视的门:检测

Teradyne 在同一天抛出了"MicroLED  测试挑战",核心论点是一句大实话:单晶圆集成器件数上升后,测试吞吐与数据分析,已经成为与转移效率并列的量产瓶颈。

这点国内讨论得很少,所有人的注意力都在"怎么把芯片放上去"。

但这里有个很容易被忽略的算术题:99.9% 的良率听起来很高,可对 1 亿颗芯片来说,0.1% 就是 10 万个坏点。

换句话说,即便转移环节做到了 99.9%,检测节拍跟不上、数据吞吐卡住,整线产出照样上不去。

巨量转移之后,检测与数据管理是规模化的第二道门槛。

对通信器件来说更是如此——它对一致性和可靠性的要求,比消费表明严苛得多。

铜,真的撞墙了

如果说工艺进展是慢变量,需要侧的爆发就是快变量。

铜已经到墙了。在 400G/800G/1.6T 这些速率下,传统铜缆信号衰减严重,高带宽下传输距离不足 2  米,还会显著发热。而机架功率密度已经突破100kW。

规模向十万卡迈进的集群里,相当比例的电力不是花在计算上,而是耗在"把数据搬来搬去"这个动作本身上。

巨头开始定调。据会议综述,英伟达在 OFC 2026 期间把 MicroLED 光互连列为下一代 AI 工厂标准接口的备选思路,定位为  CPO/NPO 的光源路线之一,与 VCSEL、硅光形成互补。

也有公开资料提到,2026 年 3 月 AMD、Broadcom、Meta、Microsoft、NVIDIA、OpenAI 等联合发起的 OCI MSA  中,MicroLED 被列为 50 米以内短距的核心光源路线之一。

产业链已经动起来了:

海外——Avicena 的 LightBundle eKit 自称是全球首个 MicroLED 光互连评估平台,现场演示  512Gbps,与台积电合作推进生产,早期投资人里还有三星和美光;联发科与微软研究院等联合展示 MicroLED 有源光缆原型;Marvell 携手 Mojo  Vision 开发高密度互连。

台系——富采具备 4/6/8 英寸晶圆制程与从磊晶到封装的垂直整合能力,被视作光源端关键角色;台积电 COUPE 思路使用 3D  异质集成。

大陆——京东方旗下华灿光电建成 6 英寸 MicroLED 芯片量产线,通信应用芯片已交付海外客户验证;三安光电联合清华大学、中国移动研制出  3dB 带宽超过 7GHz 的器件;兆驰股份在推进 400G/800G 高速光模块。

时间表上,产业界态度一致但谨慎:2026 年是落地元年(评估套件与样品),2027—2028 年进入量产爬坡,规模化应用预计在2028  年前后。

同一条产线,换一把尺子

这一段是我最想写给国内从业者看的。

首先,它给 LED 芯片厂开了"第三条路"。过去 MicroLED 的商业化叙事只有两条:电视(高端、量小)和 AR  眼镜(要求苛刻、迟迟不起量)。光互连是第三条——既不依赖电视需要回暖,也不用干等 AR 眼镜爆发。

其次,它对工艺的要求"刚刚好"。CPO 光源要的是通道数量和可靠性,不是极致的像素密度,而且允许冗余。国内已经建起来的 6 英寸  MicroLED 产线,在表明端可能还不够打 AR,但在光互连端正好够用。

产线还是那条产线,只是换了一把尺子来量。

但必须看清两个数字的量级差。机构预估,2030 年 MicroLED CPO 光收发模块产值约8.48  亿美元;美银预测同期 AI 光互连市面规模将达730 亿美元。

也就是说,MicroLED 在这个大市面里,当下只对应约1% 的份额预期。

它是切入点,不是终局。

CPO 在 AI 数据中心光通信模块中的渗透率,预计从 2026 年约 0.5% 升至 2030 年接近 35%。MicroLED  能不能吃到这份渗透率,取决于它能否先证明自己比 VCSEL 和硅光更省电,而且更好造。

先别急着高兴:三道坎还没过

好消息讲完了,坏消息也得摆出来。尤其是前面那个"成本没打平也能拿订单"的推论——它要成立,得先过三道坎。

坎一:工艺路径还没收敛

芯片尺寸、单通道速率、接收端到底用 PD 还是 CMOS,行业仍在探索。多数企业还处在外观、送样和可靠性测试阶段。

光学耦合效率与芯片良率,是公认的核心卡点。

坎二:TCO 这笔账,当下还只是纸面上的

这是最关键的一条。

"不按寸算钱"的客户愿意为省电付溢价——前提是省下来的钱真的多于价差。

而产业界普遍的口径是:2027 年才是成本拐点(打平),2028—2029 年才形成显著长处。也有人更保守,判断还需要 3~5 年。

在此之前,MicroLED 光互连是"更省电,但未必更便宜"。

而且通信客户虽然不按寸算钱,却有另一套更磨人的规矩:认证周期长、可靠性门槛高。送样到量产之间的路,可能比面板客户更难走。

那扇窗口理论上存在,但推开它需要时间,也需要良率和成本同步跟上。

坎三:表明端的窗口正在被 OLED 压缩

这是最容易被忽略的一层,而且——就发生在这三天里。

Garmin 表示,Fenix 9 那块 3000 尼特的 OLED 已经够用了,暂无推出 MicroLED 版本的计划。

三星表明的 M16 OLED,把手机屏峰值亮度推到了 10000 尼特(全屏 2800 尼特),凭第二代无偏光片工艺拿到约 30%  效率增益——而就在一年前,它才刚做到 5000 尼特。

与此同时,三星还拟投约 3000 亿韩元建 12 英寸 RGB OLEDoS 线,正面对准 XR 微表明。

MicroLED 原本最坚固的"亮度代差"长处,正在被 OLED 的快速进步一点点压缩。

如果它继续把全部赌注押在表明上,很可能面临两头挤压:电视端成本下不来,AR 端被 OLEDoS 抢先卡位。

光互连在这个时点浮现,与其说是新市面,不如说是战略缓冲。

先去数据中心“上班”,再回客厅

把线索串起来,暗线实际上很清晰:

一边,是转移、键合、检测三道工程门还没完全打通,量产线苦等了十年。

一边,是AI 数据中心的铜互连已经撞墙。

中间,是 MicroLED 那个被误读了十年的物理特性——"宽而慢"在表明里是缺点,在通信里是答案。

所以我的判断是:MicroLED 的第一个大规模量产情形,很可能不叫"表明"。

它会先以"光互连光源"的身份,被那批不按寸算钱的客户推上量产线;在那里把工艺跑熟、把成本降下来,再回头去打电视和  AR——而不是相反。

对从业者来说,这意味着评估一个 MicroLED 项目的坐标系要改一改了。

别只问"能不能做进电视、做进 AR 眼镜"。

也问一句:这套工艺,能不能先给那批按瓦算账的客户供上货?

一个难题正在从"工艺能不能做",变成"谁来先付钱"。

而这一次,掏钱的可能不是消费者,是数据中心。

你觉得 MicroLED 会先在数据中心跑通,还是先在 AR 眼镜上爆发?评论区聊聊。

本文来源:https://so.html5.qq.com/page/real/search_news?docid=70000021_4956aaf6a9578652
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

A company that makes aiming sights for the US military has started supplying GPUs

First up is a company you might not have heard of: Kopin.

It has made military microdisplays for decades—helmet-mounted displays, thermal imaging sights, and individual soldier wearable devices. In one sentence: displays for soldiers to look at.

Its current flagship product is called Neural I/O, and the customers are not soldiers but data centers.

The timeline goes like this:

In April 2026, Fabric.AI placed a $15 million initial development order with Kopin to fund the Neural I/O demonstration chipset. In return, Kopin received a 19.9% equity stake in Fabric.AI and became the exclusive manufacturer of the chipset.

On September 17, the two sides retained law firms to advance a joint intellectual property strategy.

The underlying principle, once explained, is quite simple: repurpose programmable MicroLED pixels into ultra-high-speed optical transceivers that move data with photons instead of electrons—replacing the copper wires between GPUs in AI data centers.

After a decade-long wait, MicroLED's mass production line may be snapped up first by clients who aren't charged by the inch.

The calculation model in the joint white paper shows that in a cluster of 100,000 GPUs, MicroLED optical interconnects could reduce related power consumption by about 75% compared with traditional optical approaches, for a net savings of roughly 30MW of continuous electrical demand.

What does 30MW mean? It is not just tens of millions of dollars a year in electricity and cooling costs. The more practical significance is that operators may not need to build new substations, nor delay AI projects because of grid interconnection constraints.

A display process used for night-vision sights being turned against the power wall of GPU clusters sounds a little absurd.

But it exactly exposes the awkward position MicroLED has been in for over a decade.

Waiting ten years: it is stuck at "not a single one can be wrong"

MicroLED has been called the "next-generation display technology" for more than ten years. High brightness, low power, long life, no burn-in risk—on paper it crushes LCD and OLED across the board. Samsung and LG show off new things on their booths every year, from TVs to watches to automotive displays, impressive every year.

Yet after more than a decade, it still mostly lives on the booth.

What blocks it is mass production.

MicroLED chips cannot be grown directly on a display backplane. They have to be fabricated on substrates such as sapphire or gallium arsenide at temperatures above 700°C, then transferred one by one onto the backplane.

How many? An 8K TV has 7680×4320 pixels times three colors, roughly 100 million micron-scale chips.

To put it another way: at one chip per second, 24 hours a day, it would take more than three years.

And that is only the quantity problem. The more serious issue is the second one—

one bad pixel ruins the entire screen.

That is why it has waited ten years: it is not that it cannot be made, but that it cannot be made cheaply enough or reliably enough. If the process does not run smoothly, yield does not go up and cost does not come down.

So here is the question: why would the same chip suddenly become easier in a data center?

Three flaws that are fatal in displays are not problems in communications

The answer is that the evaluation criteria are completely flipped.

MicroLED has three fatal pain points in displays, and none of them is much of a problem in communications.

First, color. Displays need RGB, and the brightness and wavelength of the three colors have to match. Color separation and alignment alone are a headache. Communications? Monochrome is enough—the mainstream approach uses parallel channels of the same color, skipping the most troublesome step entirely.

Second, bad pixels. A single dead pixel is visible and unacceptable on a display. Communications can build in redundant channels; if one fails, another takes over, and the user never notices.

Third, brightness uniformity. Mura (non-uniform brightness) is unacceptable on a display, and the cost of binning chips for sufficient uniformity is extremely high. Communications cares about bit error rate and power, not brightness uniformity.

And the performance metrics below hit the bullseye in every single category for communications:

Energy per bit of 1–2 pJ/bit, about 5% of a copper-cable approach at the same speed and distance;

Withstands 125°C and lasts over 30,000 hours. Because it is based on spontaneous emission, it is naturally immune to heat, so no TEC temperature control is needed;

Because it is a parallel low-speed architecture, it also eliminates high-power components such as DSP and CDR. Platform complexity and cost both go down.

The industry calls this approach "wide and slow"—instead of pushing a single channel faster and faster, it uses hundreds or thousands of low-speed channels in parallel, trading quantity for total bandwidth.

Microsoft's MOSAIC prototype has reportedly achieved stable single-channel 2Gbps transmission over 20 meters; Avicena demonstrated a 512Gbps link live at OFC 2026 and plans to push toward 896Gbps.

Which brings us to the most important sentence in this article:

The characteristics that are most fatal for MicroLED in displays—"full color is hard, bad pixels are fatal, uniformity is demanding"—are barely problems at all on the communications side.

The shortcoming it has been criticized for in displays for a decade is its ticket into the AI era.

To be clear: this does not mean the communications side is "easy." Integrating hundreds of channels in 1mm² works out to a pitch on the order of tens of microns, so the density requirement is still substantial.

The real relief comes from those four points—monochrome, redundancy, no color-separation alignment, and bad pixels that do not affect the viewing experience. Put together, the engineering difficulty drops off a cliff.

Panels are sold by the inch; optical interconnects are sold by the watt

Once the process story holds up, there is a more practical hurdle: who pays?

This brings us to a basic rule of the display industry.

How are panels sold? By size, by sheet. The price list says it plainly: 32 inches this much, 55 inches this much, 65 inches this much. Customers compare nothing but "price per inch."

So the only survival rule for panel makers is to keep pushing unit cost down as size and generation lines scale up. Fail to do that, and you are out.

Now switch to the data center side.

When cloud providers buy optical interconnects, the first question is not "how much does this sheet cost." They calculate a different equation: how much electricity does it take to move one bit of data? Can the cooling handle it? Do we need to build a new substation to cover the extra power?

That is what the headline means by "not priced by the inch."

This difference leads to a particularly important corollary:

To a panel customer, a MicroLED chip is a cost item—something to be squeezed on price. To a data center customer, it becomes a power-saving tool—something that saves electricity bills and substation investment.

One is squeezed on price; the other is a solution that can share in the savings.

That is why a customer who does not pay by the inch might place an order before MicroLED costs have even reached parity.

Note the "might." Most projects are still in the sample-validation stage, and the industry consensus is that the cost inflection point will not arrive until around 2027. Whether this window can actually be captured is discussed later.

Flushing chips into pits: on the process side, there was just a conference

Now let us look at the process.

On September 17, MicroLED Connect 2026 took place in Eindhoven, the Netherlands. The topics happened to answer "where exactly are yield and cost stuck?"

Route one: compress "transfer" and "bonding" into a single step

The process shown by Korea's ETRI (Electronics and Telecommunications Research Institute) is called SITRAB (simultaneous transfer and bonding).

The traditional process is two steps: transfer first, then bond. SITRAB uses a proprietary solvent-free adhesive coated extremely thin and uniform, and after a few seconds of 980nm infrared laser exposure, transfer and soldering are completed in one step.

The truly clever part is "coat once, process multiple times": after six rounds of 300W, 6-second laser irradiation, the adhesive's core functional groups remain active—it does not fully cure until above 120°C. That sounds like a materials-science detail, but it delivers three capabilities the display world has long dreamed of:

Scalable tiling. A 15×15 array can be assembled four times into a 30×30 array with seamless expansion, pointing directly at large-format panels.

Repairability. Redundant chips can be transferred to repair electrodes without removing the bad chip. The research team's data shows that redundancy repair can raise faulty-panel yield from 83.01% to 99.80%.

Heterogeneous integration. Chips of different materials, colors, and thicknesses (for example, red at 9μm, green/blue at 6μm) can be integrated sequentially onto the same backplane.

The results in the paper include a 165 PPI AlGaInP/InGaN hybrid backplane and a 32×32 full-color panel (324μm pixel pitch, 50μm subpixel pitch).

Route two: using slot-die coating to flush chips into traps

Pixel-Flo, a spin-out from the University of Sheffield, takes a different path.

Its judgment is blunt: mechanical pick-and-place has fundamental limits—square carriers do not match round wafers, so edge chips are wasted; worse, unit cost does not fall as substrate size increases. In the panel industry, that is almost a death sentence.

Continuous-Flo is a vivid concept: release known-good chips into an "ink," use standard lithography to create traps on the substrate, then use a modified slot-die process in which meniscus forces continuously guide chips into the traps. Deposition of 4μm chips at 25μm pitch has already been demonstrated.

The benefits are straightforward: edge chips can be used, smaller and cheaper chips can be used, fewer steps, higher throughput, unit cost falls with substrate size, and RGB can be done in a single step.

The company was spun out of the University of Sheffield in 2025 and completed a £5.25 million seed round (about RMB 47.8 million) in July 2026, and it specifically hired a business development lead based in Taiwan, China, to connect with the Asian panel supply chain.

TU Ilmenau and other approaches were also presented at the same event. Several routes running in parallel shows that the engineering approach has not yet converged—but it also shows the industry is no longer debating "can it be done" and is instead competing over "which one is cheaper."

A door that has been overlooked: inspection

On the same day, Teradyne raised the "MicroLED test challenge," and its core argument was a plain truth: as the number of devices per integrated wafer rises, test throughput and data analysis have become mass-production bottlenecks on par with transfer efficiency.

This is rarely discussed in China; everyone's attention is on "how to place the chips."

But there is an easy-to-miss arithmetic problem here: a 99.9% yield sounds high, yet for 100 million chips, 0.1% is 100,000 bad pixels.

In other words, even if transfer achieves 99.9%, if inspection cadence cannot keep up and data throughput gets stuck, overall line output will still fail to scale.

After mass transfer, inspection and data management are the second threshold for scaling.

This is especially true for communications devices, whose consistency and reliability requirements are far stricter than consumer displays.

Copper really has hit a wall

If process progress is a slow variable, the demand-side explosion is a fast variable.

Copper has hit the wall. At 400G/800G/1.6T rates, traditional copper cables suffer severe signal attenuation, manage less than 2 meters of transmission distance at high bandwidth, and generate significant heat. Rack power density has already exceeded 100kW.

In clusters scaling toward 100,000 cards, a significant share of electricity is spent not on computation but on the act of "moving data around."

The giants are setting the tone. According to the conference summary, during OFC 2026 NVIDIA listed MicroLED optical interconnect as a candidate standard interface for next-generation AI factories, positioning it as one of the light-source routes for CPO/NPO, complementary to VCSEL and silicon photonics.

Public materials also indicate that in the OCI MSA jointly launched in March 2026 by AMD, Broadcom, Meta, Microsoft, NVIDIA, OpenAI, and others, MicroLED was listed as one of the core light-source routes for short-reach interconnect within 50 meters.

The supply chain is already moving:

Overseas—Avicena's LightBundle eKit claims to be the world's first MicroLED optical-interconnect evaluation platform, demonstrated 512Gbps live, and is working with TSMC to advance production; early investors include Samsung and Micron. MediaTek and Microsoft Research jointly demonstrated a MicroLED active optical cable prototype. Marvell is developing high-density interconnects with Mojo Vision.

Taiwan-based players—Ennostar has 4/6/8-inch wafer processes and vertical integration from epi to packaging and is seen as a key light-source player; TSMC's COUPE approach uses 3D heterogeneous integration.

Mainland China—HC Semitek, under BOE, has built a 6-inch MicroLED chip mass-production line and delivered communications-application chips to overseas customers for validation; Sanan Optoelectronics, together with Tsinghua University and China Mobile, has developed devices with 3dB bandwidth above 7GHz; Zhejiang Jiangdong (Zhejiang Jiangdong/China Star?—correct: Jiangsu Zhaori? no, it is "Zhaochi") is advancing 400G/800G high-speed optical modules.

On the timeline, industry attitudes are consistent but cautious: 2026 is the landing year (evaluation kits and samples), 2027–2028 is production ramp-up, and large-scale application is expected around 2028.

The same production line, with a different ruler

This is the section I most want to write for domestic practitioners.

First, it opens a "third path" for LED chip makers. In the past, MicroLED's commercialization story had only two tracks: TV (high-end, low volume) and AR glasses (demanding, slow to take off). Optical interconnect is the third—it does not depend on TV demand recovering, and it does not require waiting for AR glasses to explode.

Second, its process requirements are "just right." CPO light sources need channel count and reliability, not extreme pixel density, and they allow redundancy. The 6-inch MicroLED lines already built in China may not be enough for AR on the display side, but they are exactly adequate for optical interconnect.

It is the same production line, just measured with a different ruler.

But you must see the difference in magnitude between two numbers. Institutions estimate that by 2030, MicroLED CPO optical transceiver module output will be about $848 million; Bank of America predicts the AI optical-interconnect market will reach $73 billion in the same period.

In other words, MicroLED currently accounts for only about 1% of the expected share of this large market.

It is an entry point, not the endgame.

CPO penetration in AI data center optical communication modules is expected to rise from about 0.5% in 2026 to nearly 35% by 2030. Whether MicroLED can capture this penetration depends on whether it can first prove it is more power-efficient than VCSEL and silicon photonics, and easier to manufacture.

Do not celebrate too soon: three hurdles remain

The good news is done; now the bad news. Especially the earlier corollary that "orders can be won before costs reach parity"—for that to hold, three hurdles must be cleared.

Hurdle one: the process route has not converged

Chip size, per-channel data rate, and whether the receiver uses PD or CMOS are still being explored. Most companies are still in the design, sampling, and reliability-testing stages.

Optical coupling efficiency and chip yield are recognized as the core sticking points.

Hurdle two: the TCO equation is still only on paper

This is the most critical one.

A customer who does not pay by the inch is willing to pay a premium for power savings—but only if the money saved actually exceeds the price difference.

And the industry's general view is that 2027 is the cost inflection point (parity), with a significant advantage forming only in 2028–2029. Some are more conservative, saying it will take another 3–5 years.

Before then, MicroLED optical interconnect is "more power-efficient, but not necessarily cheaper."

Moreover, communications customers may not pay by the inch, but they have another, more grinding set of rules: long certification cycles and high reliability thresholds. The road from sampling to mass production may be harder than with panel customers.

The window theoretically exists, but pushing it open takes time, and yield and cost must improve in parallel.

Hurdle three: the display-side window is being compressed by OLED

This is the easiest layer to overlook, and it has been happening in just the past three days.

Garmin says the 3,000-nit OLED in the Fenix 9 is good enough and has no current plan to release a MicroLED version.

Samsung Display's M16 OLED pushes smartphone peak brightness to 10,000 nits (2,800 nits full-screen), with roughly a 30% efficiency gain from second-generation polarizer-free technology—just a year after reaching 5,000 nits.

At the same time, Samsung reportedly plans to invest about KRW 300 billion to build a 12-inch RGB OLEDoS line, aimed squarely at XR microdisplays.

MicroLED's strongest and most durable "brightness gap" advantage is being compressed bit by bit by OLED's rapid progress.

If it keeps betting everything on displays, it may face pressure from both sides: TV costs will not come down, and the AR end will be captured by OLEDoS first.

The emergence of optical interconnect at this moment is less a new market than a strategic buffer.

First go to work in the data center, then return to the living room

Connecting the threads, the hidden line is actually clear:

On one side, the three engineering gates of transfer, bonding, and inspection are not fully open, and mass-production lines have waited ten years.

On the other side, copper interconnect in AI data centers has hit a wall.

In between is the MicroLED physical property that has been misunderstood for a decade—"wide and slow" is a weakness in displays and the answer in communications.

So my judgment is that MicroLED's first large-scale mass-production scenario is likely not called "display."

It will first go into mass production in the identity of an "optical interconnect light source," pushed by those customers who do not pay by the inch; there it will mature the process and bring down cost, and only then turn back to attack TV and AR—not the other way around.

For practitioners, this means the coordinate system for evaluating a MicroLED project needs to change.

Do not only ask "can it go into a TV or AR glasses."

Also ask: can this process first supply the customers who pay by the watt?

The question is shifting from "can the process be done" to "who pays first."

And this time, the one paying may not be the consumer—it is the data center.

Do you think MicroLED will first scale in data centers or explode in AR glasses? Share your thoughts in the comments.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry