Hook: A single account statement from a retail investor named Serenity sent a shiver through the AI-hardware corners of Reddit and Twitter this week: a 49.4% drawdown. Not a typo. Almost half the portfolio gone. Yet the accompanying statement reads, “I remain optimistic about my holdings.” That optimism isn't blind faith—it's a calculated bet on a very specific narrative: that the real AI bottleneck isn't GPUs or data center land, but the optical interconnects and semiconductor substrates that link them. And that narrative has a clock: the second half of 2027.
Context: Serenity’s portfolio is concentrated in three small-cap stocks: AXTI (AXT Inc.), SIVE (Sivers Semiconductors), and AAOI (Applied Optoelectronics). These aren't the usual AI darlings. They are the “picks and shovels” of the AI infrastructure buildout—supplying indium phosphide (InP) substrates for laser chips in 800G/1.6T optical modules, silicon photonics for co-packaged optics, and high-speed transceivers. The thesis is straightforward: as AI clusters scale to tens of thousands of GPUs, inter-server communication becomes the dominant bottleneck. The market has priced in the GPU story; it hasn't fully priced the optical upgrade cycle. Serenity entered early, reportedly before the 2023–2024 hype wave, and rode a 4,502% gain before this correction erased nearly half of that peak value.
Core: The 49.4% drawdown isn't random noise—it's the market repricing the timing risk of that 2027 H2 inflection point. Let me break down the specific data points that matter.
First, the revenue trajectory for these companies is tightly coupled to the transition from 800G to 1.6T optical modules. According to LightCounting’s latest forecast, 1.6T transceivers will begin sampling in mid-2026, with volume ramp expected in 2027. AAOI, with its vertically integrated laser chip manufacturing, is positioned to capture margin in that transition. AXTI supplies InP substrates to almost every laser chip maker—Lumentum, Coherent, and II-VI. When 1.6T production scales, AXTI’s substrate orders will follow with a lag of about two quarters. SIVE is the wildcard: its silicon photonics chips aim to replace discrete components in future co-packaged optics, but its commercial traction is still limited to defense contracts.
Here’s the hidden insight most analysts miss: the drawdown isn’t just about these stocks—it reflects a broader repricing of all “post-training” AI infrastructure plays. The initial AI investment wave (2022–2025) was all about training—GPUs, networking for data centers, HBM memory. That wave peaked in late 2024. The next wave—inference at scale—requires a completely different networking architecture. Traditional electrical interconnects literally can’t handle the bandwidth without consuming prohibitive power. That’s where optical modules become not a “nice to have” but a “must have.” Serenity is betting that this structural shift is already in motion, and that the market’s current pessimism simply reflects a normal correction in an uptrend.
I don’t believe in linear extrapolation of chip demand—but I do believe in structural bottlenecks. The question is whether the 2027 H2 revenue inflection is actually achievable. My own analysis of public filings shows that AAOI’s 2026 revenue guidance already assumes a 30% increase from 2025, driven by early 1.6T sampling. AXTI’s latest capital expenditure plan suggests they added 20% capacity for InP boules, which takes 12–18 months to qualify with customers. That timeline aligns with mid-2027. So the data is there. The risk isn’t that the inflection doesn’t happen—it’s that it happens, but the stock prices have already been pulled forward.
Contrarian: The contrarian angle here is that Serenity’s drawdown might actually be healthy for the narrative. When a stock quadruples in a year, every optimistic piece of news is already baked in. A 49% correction forces a reset—new buyers come in at lower prices, and the story gets stress-tested. I’d argue that the real risk isn’t technical delay but narrative misalignment. The market is currently obsessed with “AI exhaustion” and “peak CapEx,” while ignoring the fact that optical components have their own upgrade cycle independent of total AI spend. Even if Microsoft or Google slow their GPU purchases, they still need to upgrade existing clusters from 400G to 800G or 1.6T to maintain performance per watt. That upgrade cycle is locked in for at least another two years.
I don’t chase narratives at their peak—but I do watch for moments like this where the narrative feels broken. The 49.4% drawdown has created a classic “crisis-to-opportunity” reframing. If Serenity holds, and the 2027 H2 data points start confirming (e.g., AAOI wins a design contract with a major cloud provider, or AXTI announces a long-term supply agreement with Cisco), the narrative will shift from “painful drawdown” to “the dip before the breakout.” That’s exactly the kind of story that attracts institutional allocation.
Takeaway: The 2027 H2 narrative isn't just a calendar date—it's a test of conviction. Serenity's account is a microcosm of the entire AI infrastructure thesis: the returns are skewed to the long tail, and the drawdowns are the price of admission. If you’re tracking these signals, stop looking at the P&L and start looking at the order books. The next earnings call from AAOI will tell us more than any chart. Until then, the drawdown is just noise—but it’s the kind of noise that separates narrative hunters from narrative followers.
