Alexandr Wang's first year inside Meta can be measured in two unusually large numbers. The first is nine months, the time Meta Superintelligence Labs says it took to rebuild the company's artificial-intelligence stack and release Muse Spark. The second is $125 billion to $145 billion, Meta's forecast range for 2026 capital expenditure, including finance-lease principal payments. Wang has delivered a model. He must now help prove that the infrastructure, talent and corporate disruption behind it can produce a return.
Muse Spark arrived in April as the first model in a new series from the laboratory Wang leads as Meta's chief AI officer. It was deliberately designed to be smaller and faster than the frontier systems that absorb the industry's attention. Meta placed it in the Meta AI app and website, then began rolling it through WhatsApp, Instagram, Facebook, Messenger, Threads and its artificial-intelligence glasses. By May, the company had added faster voice responses, live visual assistance and shopping functions that draw on Marketplace, maps, brands and creators.
This is a strategically coherent answer to Meta's position. It does not need to win every academic benchmark to create value. It needs a model that works at low enough cost and high enough speed across services used daily by 3.56 billion people. It can improve recommendations and advertising, increase the utility of messaging, strengthen a standalone assistant and make glasses more useful. The distribution advantage is immense. So is the obligation to turn deployment into economics rather than another expensive demonstration.
A costly reset
Wang joined Meta in June 2025 after the group made a minority investment in Scale AI that Scale said valued the data company at more than $29 billion. Meta's annual filing records the cost of its Scale stake at $13.8 billion. Wang left the chief executive role at the company he founded and remained a Scale director, while Meta created a new centre of authority around him. The transaction offered liquidity to Scale shareholders, expanded a commercial data relationship and, most importantly for Meta, brought in an executive to lead its renewed model effort.
The structure shows how urgently Mark Zuckerberg wanted to recover ground. Meta had helped popularise open-weight models through Llama, yet rivals were setting the pace in consumer assistants, coding and frontier reasoning. Rather than rely on incremental reorganisation, Zuckerberg paid the equivalent of a major acquisition price for a minority holding and access to leadership and talent. That choice made Wang one of the most closely watched executives in technology before he had shipped a Meta product.
Muse Spark changes the burden of proof but does not remove it. Meta says the model can reason across science, mathematics and health, work with voice, text and images, and run parallel subagents for more complex tasks. It is also being offered to selected partners in a private application-programming-interface preview. Larger Muse models are in development. These are product claims, not yet a disclosed business segment. Meta does not report revenue, operating profit or usage for its generative-AI assistant separately.
The financial base gives Wang time. Meta generated $56.31 billion of revenue in the first quarter of 2026, up 33 per cent, and $22.87 billion of operating income. Advertising impressions rose 19 per cent and the average price per advertisement increased 12 per cent. Operating cash flow was $32.23 billion. The core business can support investment at a scale most laboratories cannot contemplate.
Yet the investment is rising faster than the comfort provided by those figures. First-quarter capital expenditure was $19.84 billion, and Meta lifted its full-year forecast by $10 billion at both ends, citing higher component prices and additional data-centre costs. Its 2025 capital expenditure had already reached $72.22 billion. The company also ended that year with $58.74 billion of long-term debt, double the previous year's level, even though it retained $81.59 billion of cash and marketable securities.
Distribution is not monetisation
Wang's commercial opportunity sits inside Meta's existing revenue engine. Better models can improve content ranking, help advertisers create campaigns and generate stronger signals about intent. Meta began using interactions with its AI features to personalise content and advertising in late 2025, subject to controls and exclusions for sensitive topics. Shopping results can connect queries with Marketplace listings and brand content. On glasses, a useful assistant may support device demand and create an interface less dependent on Apple or Google.
Each path carries a different test. Advertising improvements must generate incremental conversion rather than simply shifting credit from existing recommendation systems. A consumer assistant must retain users when strong alternatives are one tap away. An API must attract developers despite Meta's late entry and the uncertain boundary between its open and proprietary strategies. Glasses must become a repeatable hardware platform rather than a fashionable accessory. Lower-cost inference matters across all four.
Muse Spark's compact design is therefore more than a technical preference. At Meta's scale, a small difference in the cost of an answer becomes a large operating expense. A model that is slightly less capable but far cheaper may be the rational choice for billions of daily interactions. The larger Muse systems can handle more difficult tasks, while routing can reserve expensive computation for the moments that justify it. Wang's background at Scale, where data quality and evaluation were central products, is relevant to making that hierarchy work.
His experience is less obviously suited to some parts of the assignment. Scale built infrastructure and services for model developers, enterprises and governments; it did not operate a global consumer network funded by advertising. At Meta, model quality intersects with product design, privacy, child safety, content integrity and the competing priorities of long-established app organisations. Technical authority does not automatically resolve those institutional conflicts.
The first model already illustrates the tension. Drawing on Reels, maps, social context and Marketplace can make an assistant more personal than a generic chatbot. It can also intensify questions about which data are used, how recommendations are ranked and whether a conversational exchange becomes a commercial signal. The European Union's privacy and competition regimes place limits on the combination of data and the treatment of dominant platforms. In the United States, youth-related litigation and scrutiny of artificial-intelligence companions raise separate risks.
Asia is the scale test
Asia will reveal whether personal artificial intelligence can operate across languages, regulations and very different purchasing power. India is among Meta's largest markets and central to the reach of WhatsApp, Instagram and Facebook. Southeast Asia brings dense social-commerce behaviour and heavy messaging use. Japan and South Korea offer sophisticated consumers and device markets. Across the region, Meta can distribute a model without acquiring users one by one.
That advantage is matched by competition from local and regional systems. Chinese developers have demonstrated that capable models can be trained and served with aggressive attention to cost, while many Asian governments want sovereign infrastructure and greater control over data. Local languages are not a marginal feature: quality can vary sharply across scripts, dialects, cultural context and code-switching. A model designed around English-language benchmarks may reach billions yet feel unreliable in the conversations that matter commercially.
Infrastructure also has a geographic price. Meta is developing custom accelerators, new data-centre processors with Arm and multiple generations of silicon with Broadcom. It has supply agreements with Nvidia, AMD and cloud providers. The breadth reduces dependence on any one component but increases integration complexity. Power, water, grid access and semiconductor supply chains link the build-out to Asia even when a data centre stands in America. Export controls can restrict where the most advanced hardware is deployed and which researchers can collaborate.
Meta once made openness a central distinction of its model strategy. Muse Spark's private API preview and product-first design suggest a more controlled phase. That may protect competitive advantage and simplify safety enforcement, but it weakens the ecosystem argument that made Llama influential across Asian developers, universities and smaller companies. Wang must determine which layers remain open, which are proprietary and how the two reinforce rather than cannibalise each other.
The authority beneath Zuckerberg
Wang is chief AI officer, not chief executive. Zuckerberg sets the ambition, controls Meta through its dual-class share structure and has repeatedly shown willingness to sustain long investment cycles. That support removes one common obstacle to technical transformation. It also means the boundaries of Wang's accountability can blur. Infrastructure spending, product integration and model research involve Meta's finance, product, engineering and hardware leaders. Success will be collective; a miss could still be personalised around the expensive recruitment that created his role.
The Scale relationship adds another governance question. Meta says its investment is non-marketable and that it does not have significant influence over Scale's operations. Wang remains on Scale's board while leading a major customer and investor. Clean controls over procurement, data, staffing and competitive information are essential, particularly as other model laboratories decide whether Scale remains a neutral supplier. Meta obtained a strategic asset; it must not inadvertently reduce that asset's value by making independence less credible.
For now, the operating evidence favours patience. Muse Spark shipped quickly, reached existing products and coincided with accelerating revenue and stable operating margins. Meta expects 2026 operating income to exceed the 2025 level despite the infrastructure step-up. A first model that supports advertising and engagement could justify part of the bill even before a separate assistant business emerges.
But capital expenditure of up to $145 billion changes the standard. Wang cannot be judged only by model releases, benchmark charts or the speed of hiring. He needs to show lower inference costs, measurable product lift, reliable performance across markets and a route from personalisation to revenue that regulators will tolerate. The next Muse model may be larger. The more important milestone will be evidence that Meta's artificial intelligence is becoming economically larger than the infrastructure built to run it.
