# AI capex — X 热门讨论 (2026-09-23 21:02 UTC)

## @piovincenzo_ (Pio) · 09-23 17:59 · ♥31 ↻0 💬13 Tesla is deploying an insane amount of capital

$8.3B in capex in the first half of 2026

AI computing Data centers Optimus Robotaxi

The car business is still paying the bills

But they are clearly positioning for a much bigger business https://x.com/piovincenzo_/status/2102820103079256266

## @carrioresearch (Carrio) · 09-23 17:53 · ♥31 ↻4 💬3 $SIVE SIVERS, PIECED TOGETHER: 16 LASERS PER GPU ON NVIDIA'S ARCHITECTURE PUTS 2028 DEMAND AT 304 MILLION, AGAINST A GLASGOW LINE SIZED FOR 100 MILLION

Put everything on the public record about Sivers Semiconductors side by side and two things sit at the centre. The first is where this laser chip is meant to sell and on what schedule. The second is how far the company has since turned up the dial on its own line.

The hardest thing comes from photonics chief technology officer Andrew McKee, on co-packaged optics:

- 19.0 million AI GPU shipments forecast for 2028 - One laser source drives two silicon photonic chiplets, the typical architecture today - Ayar Labs' approach puts 10 chiplets around a GPU, which works out to five laser sources - The architecture NVIDIA has shown puts 32 chiplets around a GPU, which works out to 16 - Below 5 pJ/bit for DWDM microrings, against around 15 pJ/bit for pluggable optics including LPO

Multiply those lines out and a single year of 2028 consumes more than 95 million laser sources on the Ayar Labs count, and 304 million on NVIDIA's. The multiplication is ours; both sets of numbers came off the stage. McKee put the range into one sentence himself: "The potential market size for this is approaching hundreds of millions of laser sources per year as we go through 2028 into the future."

He gave the schedule by year as well. 2025: design frozen, technology developed, qualification starts. 2026: qualification completed, pilot ramping begins. 2027: volume production. He paused on the CW-WDM MSA slide — the consortium setting the wavelength grids and optical power requirements counts NVIDIA, AMD and Inphi, now part of Marvell, among its observer members.

The lasers themselves are prototyped at the company's own wafer fab in Glasgow, on a ridge waveguide architecture with etched facets, on-wafer optical coatings and on-wafer optical test.

Mark Wade, founder and chief executive of Ayar Labs, appears in person at the session, and what he talks about is capacity. What he says he is excited about is Sivers taking the core innovations accomplished in its research manufacturing environment — the Glasgow fab — and transferring those capabilities into a true high-volume, commercial-ready production foundry environment, which is what it is doing at WIN Semiconductors. On the scale the industry will need, he left a line: "tens of millions, maybe hundreds of millions of units per year."

Manufacturing is presented by Alex McCann, managing director of the photonics business. He opens on the early-1980s mantra that real men needed fabs, and then sets out why the company went the other way: under a foundry model the capex for capacity scale-up is carried by the foundry, and going from R&D-level volumes to tens or hundreds of millions of units is the difference between tens of millions and hundreds of millions of dollars in CapEx savings. As of May 2025 Glasgow does design, new product and process development, pre-production release and low-volume early-stage production, while volume goes to WIN Semiconductors (3105 TT) in Taiwan. WIN gets a page of its own — founded 1999, the largest pure-play compound semiconductor foundry in the world, four 150mm fabs, one of them dedicated to optoelectronics and photonics, 2024 revenue of $540 million. McCann tells the room its major customers today are Apple, Broadcom and Lumentum, and that it has the ability to build in Europe or the United States should the need arise. The wireless leg runs on GlobalFoundries' ($GFS) 45RFSOI and 22FDX nodes.

Sixteen months on, that manufacturing position has moved a step. On September 3 the company announced a USD 30 million expansion of its indium phosphide facility in Glasgow, sized on completion to support annual production of more than 100 million CW DFB lasers, with new process capabilities and more automation alongside. Work begins in the second half of 2026 and the line is expected to be operational in the fourth quarter of 2027 — landing on the same year the stage gave for volume production. The announcement calls this a transition from a Fab-Lite model to a Hybrid Manufacturing strategy: more in-house capacity, foundry, packaging and manufacturing partners kept in place.

"The scale of today's AI infrastructure build-out is creating unprecedented demand across the optical supply chain," president and chief executive Vickram Vathulya said in that announcement. "Our customers need significantly more laser production capacity and, just as importantly, the confidence that supply will be there when their programs ramp." Chief operating officer Neeraj Chopra added the manufacturing half: scaling photonics for AI infrastructure takes more than adding capacity, it takes a strategy built for flexibility, quality and long-term customer demand.

On stage in 2025 Vathulya gave a longer frame: a serviceable market above $2 billion from 2028 onwards, and what he called a potential to triple or quadruple the business over the next four to five years at very healthy gross margins. Asked directly for 2026 and 2027 revenue expectations, he said the company does not give forward-looking guidance of that kind, and that the slide was his own read on the potential, based on team size, the number of projects the company can take and the number of products it can release, rather than a company figure.

The foundry end is running warm too. WIN Semiconductors booked consolidated revenue of TWD 1.94 billion in August, up 30.55 percent from a year earlier and its highest month since the start of 2024. Its first eight months come to TWD 13.68 billion, closing on the TWD 17.46 billion it did across all of 2024. That fab runs far more than photonics, but it is the link that has to hold for Sivers to get lasers to volume.

The stage counted 304 million laser sources on NVIDIA's architecture and more than 95 million on Ayar Labs'. The line announced three weeks ago is sized above 100 million. https://x.com/carrioresearch/status/2102818702085501220

## @youngstockuser (美股投资young) · 09-23 15:26 · ♥31 ↻1 💬7 今天MAG7里面，$GOOGL这根跌得挺狠。

截至美股上午，$GOOGL大约跌3%，从昨天351美元附近直接打到340美元左右。这里并没有突然爆出Google业绩暴雷，今天压它的第一层还是利率：美国PMI远超预期以后，10Y直接冲到5.05%，Nasdaq整体跌超1%，长久期科技股全部在吃估值压力。

第二层是最近Meta Muse太火，市场开始重新讨论Google在AI Agent里的位置。过去几个月$GOOGL本身也一直被Gemini进度、AI投入太大、竞争加剧这些东西压着，所以资金一到Risk-off，Google很容易被先卖。

但如果只看公司现在实际赚的钱，这家公司没有市场情绪表现出来那么差。

Google Q2营收1198亿美元，同比增长24%；Search收入增长17%，YouTube广告增长13%，Google Cloud直接增长82%到247.7亿美元。更猛的是Cloud营业利润从去年同期28亿美元涨到88亿美元。Gemini月活已经做到9.5亿。

这就是为什么Google跌下来以后我一直觉得比很多高估值AI股舒服。

它不是靠一个未来故事撑估值。Search还在增长，YouTube继续赚钱，Cloud已经从烧钱业务变成利润机器，Gemini、TPU、Android、Chrome又全部在同一个生态里面。

当然风险也很清楚：Q2 CapEx已经高达449亿美元，如果AI投入越来越大，但Search增长开始掉、Cloud利润跟不上，那估值照样会被继续压。

但现在股价340美元左右，已经比52周高点408美元回撤了接近17%。

所以今天Google跌3%，我反而不会因为一根大阴线把它删掉。

这种现金流、Search、Cloud、AI全部还在增长的公司，每次因为利率和市场情绪被砸下来，才是重新算价格的时候。 https://x.com/youngstockuser/status/2102781778557862267