Meta (META) made two moves last week that Wall Street has been watching for. One is about cutting what it costs to run AI. The other is about charging users for it.
Together they give investors a more concrete answer to the question that has been hanging over the stock all year. Meta has been spending at a scale that makes even other hyperscalers look cautious. The question has always been when that spending starts paying back. These two moves are the first real answer.
In a note shared with TheStreet, Bank of America reiterated its Buy rating on Meta and kept its $810 price target unchanged. That implies roughly 20% upside from the stock’s September 16 close of $673.31.
Meta’s chip strategy and what it means for costs
Meta confirmed it will start deploying its third-generation custom AI chips, called Arke, in the first half of 2027. A fourth-generation chip called Astrid follows in late 2027. Both were built with Broadcom for AI inference, the process where trained models generate responses, Bloomberg reported.
Meta’s vice president of engineering told Bloomberg the chips are engineered to outperform “whatever Nvidia is currently shipping” on a per-watt and per-dollar basis. More than one gigawatt of custom chip capacity is planned over the next 12 months.
The company also tried to build a chip that could handle both training and inference. It canceled that project after finding the design would cost roughly 30% more. Not worth it.
More Meta:
- Meta sent a warning to its glasses pranksters
- Meta layoffs take disturbing turn in new lawsuit
- Meta business model in trouble from $1.4 trillion lawsuit
The note said Bank of America ran the numbers on what the savings could look like. If Meta deploys five to six gigawatts of owned capacity in 2027 at $200 billion in total infrastructure spend, with chips making up 60% of that cost, custom chips at 40% savings versus third-party alternatives could mean roughly $8.5 billion saved. That is a model estimate, not company guidance.
Google’s TPU program is worth noting here. Custom silicon took years to pay off for Alphabet but is now central to how it manages AI infrastructure costs. Meta is making the same long-term bet. The chips arriving in 2027 are the first real test of whether the economics actually hold.
Broadcom CEO Hock Tan said it directly on his own earnings call. “When you co-develop a chip that is optimized for your particular LLM workloads, you will outperform any GPU,” he said, adding that customers can do it “at half the cost.” He confirmed Meta’s program is on track and that three MTIA generations will ship by end of 2027.
What Meta One is and why it matters
Meta launched Meta One on September 15. It is a global subscription service bundling AI features, customization tools and creator capabilities across Instagram, Facebook, WhatsApp and Meta AI. More than 50 features launched with it. The service has already recorded 15 million subscriptions and trials.
Consumer plans start at $2.99 per month for individual apps and go up to $19.99 for the Premium bundle. Creator and business plans run from $14.99 to $499 per month.
The 15 million figure mixes trials and paying subscribers. Nobody knows yet how many of those convert to actual recurring revenue. But it is the first signal that people will pay for features on platforms they have used for free.
The math from the note is simple. Every 1% of Meta’s 3.6 billion users that subscribes at $10 per month average revenue adds roughly $4.3 billion a year. That is about 1.2% upside to 2028 revenue estimates. Snapchat Plus has reached 5.5% daily active user penetration. That is the benchmark Meta is chasing.
Two catalysts still ahead
Bank of America flagged five catalysts for Meta at the start of its coverage. Three are now announced: Muse, custom chips and Meta One.
Two remain. The Connect conference is first. No major AI announcements are expected there, though any Muse adoption update would be useful. Muse is Meta’s AI image and video generation system, and how quickly it is pulling users into paid tiers matters more than most people have focused on.
The bigger remaining catalyst is Watermelon, Meta’s upcoming frontier LLM. A strong model could improve Meta’s advertising systems and open API licensing fees as a new revenue line. It would also be the clearest signal yet that Meta’s AI research is producing results that translate into financial outcomes, not just benchmark scores.
What the risks look like from here
The chip savings are a model. Meta One conversion is unknown. Watermelon has not shipped. All three are bets, not bookings.
Meta still earns the vast majority of its revenue from digital advertising. A macro downturn or a pullback in ad spending hits the whole thesis. The company’s fixed asset base has also grown fast enough that cost flexibility in a downturn is more limited than it used to be.
There is also a competition risk that lies beneath the advertising risk. AI-native platforms are starting to compete for the same user attention Meta’s apps have dominated for years. If engagement shifts, the ad revenue picture changes before any of the new revenue lines have scaled enough to compensate.
The $810 target is 24 times the 2027 GAAP earnings estimate. That is a premium to the broader market. Whether Meta earns it depends on whether the chip economics land, the subscription business scales and Watermelon actually performs. None of that is settled yet.
Related: Mark Cuban exposes a problem with how Meta and Google fund AI
