Nvidia's buyback binge looks smart until you check the

The board just authorized another $50 billion in repurchases on top of the $7.7 billion they already blew through last quarter. At first glance it screams confidence — cash flow is obscene, the H100 backlog is still measured in quarters, and Jensen can honestly say "we're returning capital to shareholders" with a straight face.

But the math gets weird when you compare the yield on that buyback versus what the same dollars could earn deployed internally. Nvidia's weighted average cost of capital sits around 8.5 %. The implied yield on a buyback at today's multiple is roughly the earnings yield — call it 3.5 % after the recent run-up. That's a negative spread of five full points every year the shares don't rerate dramatically.

Meanwhile, the company is turning away sovereign AI deals because they can't allocate enough Blackwell wafers. TSMC's CoWoS-L capacity is the real constraint, not demand. Every billion diverted to buybacks is roughly 1,500 fewer B200 systems shipping in 2025 — systems that carry 70 %+ gross margins and lock customers into CUDA for a decade.

The counter-argument is that Nvidia's R&D pipeline is already saturated; they're hiring aggressively but talent absorption has limits. Fair. Yet the venture arm (NVentures) deployed only $872 million last year across 39 deals — pocket change. They could 10x that, fund the next generation of CUDA-native startups, and effectively subsidize their own moat for pennies on the dollar.

There's also the strategic signaling problem. Aggressive buybacks at all-time highs tell the market "we see no better use for cash." For a company whose entire thesis is "AI compute demand is infinite and we're the only gate," that's a contradictory message. AMD's MI300X is finally shipping volume. Intel's Gaudi 3 is sampling. Custom silicon from Google, Amazon, and Microsoft is taping out. The moat is widening, but the pace of widening depends on capital intensity — exactly what buybacks reduce.

Jensen owns 3.5 % of the float. He doesn't need the EPS boost from share count reduction. The board's comp structure is heavily equity-based, so they do. That misalignment isn't illegal, but it's worth noting when the capital allocation decision looks suboptimal on a pure IRR basis.

If the stock pulls back 20 % — not unheard of for a cyclical semiconductor name — the buyback looks brilliant in hindsight. If it compounds at 15 % for three years while capacity constraints cap revenue growth at 25 %, the opportunity cost compounds too. Right now the market is pricing perfection. The buyback assumes perfection persists. History suggests it rarely does.