Meta AI hit 1 billion monthly active users in Q1 2025, the fastest AI platform growth in history at the time. But fourteen months later, the ground has shifted. ChatGPT crossed 1 billion monthly active users of its own in June 2026, Google’s Gemini app passed 900 million monthly actives in May 2026, and Meta hasn’t published a fresh Meta AI user number since that original milestone.
The bigger story since our last update: Meta finally shipped its long-delayed flagship model, now called Muse Spark (internally, Avocado), after three deadline slips. Meta also expanded its Nvidia partnership into a multi-year, multi-gigawatt chip deal, and on July 1, 2026, it stunned Wall Street by announcing a new cloud business, Meta Compute, to sell its excess AI infrastructure. Chip stocks tumbled and Meta’s own share price climbed above $600 for the first time.
Meta isn’t just spending on AI anymore. It’s trying to turn that spending into a second business.
1 Billion Users: Still Meta’s Headline Number, But Rivals Are Catching Up
Meta AI hit 1 billion monthly active users in Q1 2025, CEO Mark Zuckerberg announced at Meta’s May 2025 shareholder meeting, at the time the fastest an AI platform had reached that scale.
Growth timeline (unchanged since the milestone):
- September 2023: Beta launch (limited markets)
- Late 2024: 500M monthly active users
- Q1 2025: 700M monthly active users
- May 2025: Crossed 1B monthly active users
- Since then: Meta has not disclosed an updated Meta AI MAU figure
Why that gap matters now: In June 2026, ChatGPT’s app crossed 1 billion global monthly active users too, according to Sensor Tower estimates reported by Reuters, the fastest app in history to reach that scale on its own metric. Google’s Gemini app, meanwhile, grew from roughly 400 million monthly users a year earlier to over 900 million by Google I/O in May 2026, and Google says AI Mode inside Search now reaches more than 1 billion monthly users on its own. Anthropic’s Claude, while still far smaller in absolute users, nearly tripled its global web traffic share in a single quarter (from about 2.2% in December 2025 to roughly 6% by March 2026).
Meta’s distribution advantage, AI built into WhatsApp, Instagram, Facebook, and Messenger, hasn’t disappeared. But the “fastest growing AI platform” story now has three or four contenders, not one.
Platform breakdown (last disclosed figures, Meta has not refreshed these):
- WhatsApp: ~630M Meta AI users, ~63% of interactions
- Instagram: ~270M Meta AI users, ~27% of interactions
- Facebook: ~100M Meta AI users, ~10% of interactions
- Standalone Meta AI app: 10-20M users, expanding beyond its original US/Canada/Australia/New Zealand markets
Geographic distribution: India still leads globally, driven by WhatsApp’s 500M+ Indian user base. Global reach sits at 60+ countries and 13+ languages.
Muse Spark Has Launched: Avocado’s Public Debut, Three Delays Later
The model long known internally as Avocado didn’t arrive on the original late-2025 timeline, or the March 2026 revision, or even the May window. It finally launched publicly on April 8, 2026, under a new name: Muse Spark, the first model out of Meta Superintelligence Labs (MSL) under chief AI officer Alexandr Wang.
What shipped:
- Muse Spark now powers the standalone Meta AI app and meta.ai, with rollout planned across Facebook, Instagram, and WhatsApp
- It accepts voice, text, and image input, but produces text only output. Despite Meta’s earlier “multimodal” framing, this version doesn’t generate images or video
- It runs in tiered modes: an instant mode for casual queries and a thinking mode for step by step reasoning tasks
- A standout feature is “visual coding,” generating interactive websites, dashboards, and mini-games from a single prompt
- It can spin up multiple sub-agents in parallel (for example, splitting a trip-planning request across itinerary, destination, and activity agents)
- All versions are free to use, with Meta planning to release an open-source variant alongside the proprietary one
How it performs: A Meta executive told Axios the model isn’t a new state of the art but is competitive with leading labs on tasks like multimodal understanding and health-related queries, scoring 89.5% on GPQA Diamond and leading rivals on HealthBench Hard at 42.8%. Meta has openly acknowledged a persisting gap in coding performance against models like GPT-5.4.
Why the three delays happened: Internal testing through Q1 2026 reportedly placed Avocado’s benchmark performance somewhere between Google’s Gemini 2.5 and Gemini 3.0, solid progress over Llama 4, but short of the frontier bar Meta needed to hit. Meta even explored temporarily licensing Google’s Gemini to power its own products while Avocado caught up, though no such deal materialized.
Mango: still unreleased. Unlike Avocado/Muse Spark, Meta’s image and video model codenamed Mango has not had a public launch as of this update. It remains in development inside MSL, aimed at competing with OpenAI’s Sora and Google’s Veo, with no confirmed release date.
Nvidia Mega-Deal: Now Public, Still Undisclosed in Dollar Terms
Meta’s expanded Nvidia partnership, announced February 17, 2026, is now fully underway.
What’s confirmed:
- Millions of Blackwell and upcoming Rubin GPUs
- Standalone Grace CPUs, Nvidia’s first large-scale standalone Grace deployment globally
- Vera CPUs slated for 2027
- Spectrum-X Ethernet networking and security tooling for WhatsApp AI features
Deal size: Neither company has disclosed the dollar figure. Analyst back-of-envelope math (roughly $3.5M per GPU-equipped rack) puts a million-GPU tranche in the ballpark of $48 billion, consistent with earlier “tens of billions” estimates.
Multi-vendor hedge continues: Meta is also evaluating Google TPUs for 2027 deployment alongside AWS Trainium, AMD GPUs, and in-house silicon, though Nvidia remains its primary supplier.
Data Centers: Hyperion and Prometheus Move From Blueprint to Buildout
Meta’s flagship sites are no longer just announcements. They’re under active construction, with clearer specs than we had in February.
- Prometheus (New Albany, Ohio): 1-gigawatt facility, tied directly to the new Nvidia chip allocation; reports point to a Meta-linked land purchase nearby suggesting further campus expansion
- Hyperion (Richland Parish, Louisiana): Now specified at 1.5 gigawatts by late 2027, scaling to 5 gigawatts by 2030, designed to house 1.3 million+ GPUs. Powered by a mix of natural gas, renewables, battery storage, and nuclear baseload, with direct liquid cooling for high-density racks
- March 27, 2026: Entergy Louisiana signed a new agreement tied to Hyperion, promising customers an additional $2 billion in savings over 20 years, a response to local concern about the data center’s power draw
2026 capex guidance: Still $115-135 billion, as set on the January 2026 earnings call. Actual Q1 2026 capex came in lower than Wall Street expected, at $19.84 billion against a $27.57 billion estimate, a gap Meta hasn’t fully explained but one that briefly rattled investors around the earnings release.
New: Meta Compute, Meta Enters the Cloud Business
The biggest development since our last update, and one that wasn’t on anyone’s radar in February: on July 1, 2026, Bloomberg reported Meta is building Meta Compute, a cloud infrastructure unit that will sell spare AI computing power to outside customers, putting Meta in direct competition with AWS, Microsoft Azure, and Google Cloud.
Why now: Meta has spent years defending its AI capex by tying it to advertising gains. But free cash flow pressure has grown alongside that spending, with capex going from $37.2B in 2024 to $69.6B in 2025 and toward $135B at the midpoint for 2026. Selling excess capacity reframes years of “unexplained” spending as a hedge, not just a bet.
Two possible shapes, according to Bloomberg’s reporting: a “bare metal” model where customers bring their own software (similar to how SpaceX rents capacity to firms like Anthropic), or a fuller cloud platform with developer tools layered on top, closer to AWS Bedrock.
Market reaction was immediate and sharp:
- Meta’s own stock jumped nearly 9%, pushing shares above $600
- Micron fell more than 10%; SanDisk dropped over 14%; Intel and AMD each lost 7-10%
- Nvidia dipped only slightly (about 1.25%), standing out against the broader rout
- CoreWeave and Nebius, both GPU-rental specialists, fell 14% and 17% respectively on fears Meta would undercut their pricing
- The sell-off spread to Asian chipmakers, with Samsung and SK Hynix both down more than 7%
The signal: for the first time, a hyperscaler openly admitted it has more AI compute than it needs right now, flipping years of assumed scarcity into a supply question. Whether Meta can actually run a cloud business (support, reliability, enterprise sales) as well as it runs social platforms is untested.
AI Revenue: Q1 2026 Numbers Are In
Meta’s Q1 2026 earnings, reported April 29, 2026, gave the clearest read yet on whether AI spending is showing up in the business.
- Revenue: $56.31 billion, up 33% year over year, Meta’s fastest quarterly growth since 2021
- Net income: $26.77 billion
- Ad revenue: $55.02 billion, 98% of total revenue, still Meta’s overwhelming monoculture
- Ad impressions: +19% YoY; average price per ad: +12% YoY, both credited to AI tooling across the ad stack
- Advertisers using Meta’s AI creative tools: 8 million, double the figure from a year earlier
- Daily Active People (DAP): 3.56 billion, a 4% YoY increase but a miss against the 3.62 billion Wall Street expected. Meta attributed the shortfall partly to internet disruptions in Iran and a WhatsApp restriction in Russia
- Click-to-message ads grew more than 50% YoY in the US; paid WhatsApp messaging kept scaling past a $2 billion annual run-rate
Takeaway: AI still hasn’t created a clean new revenue line at Meta, no “Muse Spark subscription” line item exists yet. But it’s visibly lifting the core ads engine, and Meta Compute is now a credible second bet.
Competition: The Race Has Gotten More Crowded, Not Less
OpenAI (ChatGPT):
- Crossed 1 billion monthly active app users in June 2026 (Sensor Tower/Reuters), the fastest app ever to reach that scale
- 900 million weekly active users as of February 2026
- Annualized revenue run rate over $25 billion (February 2026)
- Web traffic share has fallen sharply even as raw usage grows, from roughly 87% in January 2025 to under 57% by March 2026, as competitors take share
Google (Gemini):
- Gemini app grew from ~400M to over 900M monthly active users between May 2025 and May 2026 (Google I/O disclosure)
- AI Mode in Search now claims over 1 billion monthly users on its own
- Web traffic share climbed past 25%, more than tripling in a year
- A reported Apple deal, a custom 1.2-trillion-parameter Gemini model powering next-gen Siri, remains on track for a 2026 rollout with iOS 27
Anthropic (Claude):
- Smaller in absolute users but the fastest growing by percentage: web traffic share roughly tripled in a single quarter, from ~2.2% to ~6% between December 2025 and March 2026
- Has continued shipping frontier models through 2026, keeping pace with Meta’s own Muse Spark on several benchmarks
Where Meta still leads:
- Distribution: no rival can embed AI across apps used by roughly half the world’s population
- Ad monetization proof points: 8 million advertisers now using AI creative tools, real revenue impact visible in Q1 numbers
- Infrastructure scale: among the largest capex commitments in tech, now paired with a plan (Meta Compute) to monetize the excess
Where Meta still trails:
- Model performance: Muse Spark closed some of the gap but still lags on coding versus GPT-5.4 and top Gemini variants
- Visual generation: Mango, Meta’s answer to Sora and Veo, still hasn’t shipped
- User transparency: Meta hasn’t updated its headline Meta AI user count since May 2025, even as ChatGPT and Gemini both report fresh, larger numbers
- Enterprise trust: privacy history continues to limit adoption for sensitive workloads compared to Claude and ChatGPT
The Bottom Line
Meta’s AI bet is now playing out on three fronts at once: consumer models, infrastructure, and, as of July, a cloud business built to monetize its own excess capacity.
Results so far:
- Muse Spark live and powering the Meta AI app, with an open-source variant planned
- $56.3B in Q1 2026 revenue, up 33% year over year, with AI tooling visibly lifting ad performance
- 8 million advertisers now using Meta’s AI creative tools
- Prometheus and Hyperion moving from blueprints to active construction, backed by a firm Nvidia supply commitment
The risks:
- $115-135B in annual spending with a still-unclear return timeline
- Mango’s absence leaves a real gap against Sora and Veo in the fastest growing part of generative AI
- Meta hasn’t refreshed its headline Meta AI user numbers in over a year, while ChatGPT and Gemini both report bigger, fresher figures
- Meta Compute is unproven; running a cloud service business is a different discipline than running social platforms
- Coding performance remains a weak spot even in Meta’s newest model
Meta is no longer just trying to catch up on model quality. It’s trying to build a business model around the infrastructure it already owns, while still finishing the model race it started. Whether Muse Spark’s next iteration and Mango’s eventual launch close the remaining technical gaps, and whether Meta Compute becomes a durable revenue stream rather than a headline, will shape how the rest of 2026 plays out.
The AI race isn’t over. Meta just added a second front to it.