How much does Nvidia's $3.5B convert dilute MediaTek's EPS for existing holders?

The strangest fact about Monday's Nvidia-MediaTek announcement is not the $3.5 billion. It is that MediaTek — which reported roughly NT$185 billion, about $5.8 billion, of cash at the end of June — just took "$3.5 billion of financing" from its marquee AI partner. A company that flush does not have a money problem. It has an upside-sharing problem. And the convertible bond Nvidia bought is precisely that: a loan wrapped around a call option, with the one term that matters most — the price at which the debt turns into shares — left out of the public announcement entirely. Nvidia confirmed the amount and the partnership, and disclosed nothing else.

So here is the whole argument in one sentence before the machinery: $3.5 billion is about 1.8% of MediaTek's market value and about 0.07% of Nvidia's $5.2 trillion. This is not a financing event. It is Nvidia buying a downside-protected stake in the AI business that its own platform deals — MediaTek adopting Nvidia's NVLink Fusion and putting its N1X CPU inside Nvidia's RTX Spark PC chip — help create. Nvidia holds a bond if MediaTek falters and shares if it soars. By construction, Nvidia only converts if the custom-AI story worked, which means the dilution is self-selecting on success. Existing holders would be giving up a slice of exactly the upside the cash is meant to buy.

Now the part we can actually compute. Start from what's disclosed: MediaTek has about 1.6 billion shares out, and the market values it at roughly $198 billion total, about $124 a share. The $3.5 billion instrument, if it is ever settled in new shares, becomes a fixed number of shares set by the conversion price, which is the undisclosed term. So run the bracket, assuming a conversion price set at today's level and then at premiums to it:

• Converted at today's price: about 28 million new shares, a ~1.8% higher share count, EPS down roughly 1.7%.
• At a 30% premium to today: about 22 million new shares, EPS down roughly 1.3%.
• At a 50% premium: about 19 million new shares, EPS down roughly 1.2%.

On MediaTek's 2025 earnings of NT$66.16 per share, the worst case costs about NT$1.15 a share. To even reach 4% dilution, the conversion price would have to be set well below the market — which would itself be the story, a giveaway to Nvidia rather than an investment. The disclosed share count and market value cap this thing regardless of what the premium turns out to be. That is the whole answer to the headline question: the convertible, at its absolute worst, costs existing holders about 1.7% of earnings.

Then the offset, which decides whether the thesis survives. To put that 1.7% back, MediaTek needs roughly NT$1.8 billion (about $55-60 million) of extra annual net profit. The AI business it is spending toward is penciled at more than $2 billion of data-center revenue this year, a figure already doubled since April from an earlier $1 billion projection, and management now chases up to 20% of an ~$80 billion custom-AI-chip market in 2027. Grow that >$2 billion line by a third to a half — to roughly $2.6-3.2 billion — and at a thin 5-10% net margin the incremental profit lands somewhere in the $30-120 million range, with the $55-60 million needed to erase the worst-case hit sitting inside it. The timing lines up, too: the first custom AI accelerator is set to enter volume production in the fourth quarter. The ramp outruns the dilution, comfortably, under every conversion assumption short of a deliberate gift to Nvidia. As framed — dilution eats the growth — the thesis dies.

Which is not the same as saying the deal is free. Three things matter more than the 1.7%, and the first is genuinely unresolved: the term sheet is not public. When MediaTek files the placement, the two lines that settle everything are the conversion premium (the gap between today's ~$124 share price and the conversion price) and the settlement method. If the bonds settle in cash rather than new shares, the share count never moves and the whole dilution question evaporates. If they settle in shares at a skinny premium, the 1.7% is the cost. Read the filing, not the chatter.

Second, the margins. Custom ASIC work is priced thinner than premium smartphone silicon, and MediaTek's gross margin is already sliding — 46.2% in the June quarter, down almost three points year over year. AI revenue at razor margins pads the top line without doing much for per-share earnings; at some level it can be a headwind. Third, the scrutineers. This convert is one small slice of a pattern that has journalists asking whether Nvidia is financing the very ecosystem that buys Nvidia — Reuters flagged exactly that on this deal, and Nvidia has put an estimated $48.6 billion into ecosystem companies over the past year. For a huge, cash-generative company that is less "securities fraud" and more "supplier rents its own demand," but the honest question for holders is whether the $2 billion of AI revenue is profit MediaTek earns or demand Nvidia helped manufacture.

Step back, and here is the cheap way to hold the whole thing: a convertible is two securities in a trench coat — a bond and a call option — and the only genuinely interesting number is the boundary where it stops being the first and becomes the second. For MediaTek, that number is unfiled, and the bracket suggests it does not matter very much either way. Worry about the gross margin line instead, because that is where the AI ramp actually shows up in per-share earnings. The conversion is the part of the deal that only happens if the story works. The story is the part you actually own.