News vs. Noise: Markets Demand Proof on Rates and AI

Friday’s weak jobs report undercut the case for more rate hikes

The job report released on Friday, August 7, was a big real rates story. Nonfarm payrolls fell by 23,000 in July — economists had expected a gain of roughly 80,000 — and June’s already-soft print was revised down further, to just 20,000 jobs added from a previously reported 57,000 [15]. The unemployment rate ticked down to 4.1%, but only because fewer people were working or looking for work, and wage growth slowed to 3.2% year-over-year, the weakest pace since May 2021 [15].

That’s a meaningfully different signal than the one markets were reacting to the week before, when the Federal Reserve held rates steady on a 9-3 vote — three governors publicly pushed for a hike — and bond investors pushed the 30-year Treasury yield to a 19-year high on inflation-credibility concerns. Friday’s weak jobs data cut against that hawkish case: Treasury yields fell across the curve last week, with the 10-year down roughly 9 basis points — a basis point is one-hundredth of a percent — and the 2-year down a similar amount [16], and the market’s odds of a September rate hike dropped to around 42%, from 58% just a day earlier [17].

We’re watching this as the first real test of which story wins — sticky inflation that argues for higher rates, or a softening labor market that argues for patience. It’s exactly the kind of shift that highlights the need to stay diversified across the yield curve, rather than positioning for either outcome to resolve quickly.

SpaceX’s cash deficit is the clearest read yet on how the market is grading AI capex

SpaceX’s second-quarter report, released Tuesday, August 4, drew a sharp line between AI spending and AI cash discipline. Revenue jumped 92% year-over-year to $7.8 billion, comfortably beating expectations [18] — but quarterly capital spending came in at $18.4 billion, more than double the prior quarter’s pace, and the company ran a roughly $25 billion cash deficit through the first half of the year, meaning it spent $25 billion more than it brought in from operations [18].

The market’s reaction was immediate: the stock fell by double digits the next day, Wednesday, August 5, compounded by a share lockup expiration — the end of the window that keeps early investors from selling right after an IPO — that added a fresh supply of stock investors could sell [18], [19]. By the company’s own numbers, it’s spending roughly $6 on AI infrastructure for every $1 of AI revenue generated today — a ratio investors are increasingly unwilling to fund on faith alone [20], [18].

Given how concentrated broad equity indices have become in the handful of companies driving AI infrastructure spending, we’re watching the gap between capital spending and cash returns closely across the sector, and it reinforces our preference for diversification over concentrated exposure to any single leg of the AI buildout.

A U.S. trade rule aimed at Chinese optical components exposed a real bottleneck in AI infrastructure

A draft rule reportedly under consideration at the FCC would restrict new imports of Chinese-made optical transceivers — the small modules that convert electrical data into light so it can travel through the fiber-optic cabling inside AI data centers [5]. The policy detail matters less than what it revealed: AI clusters are shifting from copper to optical interconnects at 800-gigabit and 1.6-terabit speeds because copper can’t carry data that fast without losing signal strength or drawing excessive power, making these modules a genuine bottleneck — no interconnect, no functioning cluster [5].

China currently supplies roughly 60% of global optical-component revenue and about 56% of the contract-manufacturing capacity that assembles and tests these modules, according to research firm TrendForce [11]. That’s a meaningful concentration risk heading into a market TrendForce expects to roughly double, from $16.5 billion in 2025 to about $26 billion in 2026 [11]. McKinsey has separately estimated that 800-gigabit transceiver production could run 40%–60% short of demand through 2027 [12]. Western suppliers such as Coherent Corp, a maker of optical networking components, and Lumentum Holdings, another optical-component manufacturer, are each backed by $2 billion in Nvidia capital and multiyear purchase commitments tied to this scarcity [5].

Whether or not this specific rule survives in its current form, the underlying scarcity is real and unlikely to resolve quickly. We’re watching optical-component supply as one of the more durable physical bottlenecks in the AI buildout, and it reinforces why we think about AI infrastructure exposure through diversified, not concentrated, positioning.

Friday’s “China optical ban” headline is ahead of the facts — the scarcity above is real, the ban itself isn’t confirmed

The same draft FCC rule behind Friday’s optical-component story, above, got covered in some places as a done-deal “ban” on Chinese-made transceivers. That’s ahead of the facts: the rule is still in draft form, hasn’t been made public in detail, and could be narrowed, delayed, or dropped entirely before it takes effect [5].

Even if it is enacted, analysts covering the space have pointed out that the near-term impact would likely be limited to new transceiver models requiring approval, not existing clusters going offline, and some have questioned the security rationale itself — these modules have no external management connection or onboard storage, unlike the networking gear involved in prior China-related restrictions [5], [13]. It’s also unclear whether any rule would follow company ownership or country of final assembly, a distinction that would determine how easily Chinese suppliers could route around it by assembling elsewhere [5], [11], [13].

We’re treating the “ban” headline as a sentiment amplifier, not a confirmed policy outcome. The structural scarcity in optical components is real regardless of how this specific rule shakes out — which is exactly why we’re watching the underlying supply constraint rather than reacting to a single policy headline about it.

SpaceX’s multi-gigawatt compute targets are the company’s own forecasts, not a verified fact

Last week, SpaceX laid out ambitious AI-computing targets: more than 2 gigawatts of capacity by the end of 2026, a path toward 10 gigawatts by the end of 2027, and a longer-run pipeline of 20 gigawatts [4]. Musk himself flagged that last figure as “aspirational,” adding that the company will probably fall short of it on the stated schedule [4].

That’s worth flagging because it’s a pattern we saw more than once last week: a company describing its own future capacity, or the scarcity of the market it operates in, in terms that happen to support the stock. That doesn’t make the numbers wrong — SpaceX may well build toward those targets — but a forecast from the company itself isn’t the same as independently verified demand or shipped, revenue-generating capacity.

We treat management’s own targets as useful color on where a company thinks it’s headed, not as a number we’d size a position around. We’ll be watching for these figures to show up as actual installed capacity and cash flow before we treat them as confirmed rather than aspirational.