Google and Meta AI strategy in 2026: own chips, own models, closed or open

Google and Meta both guide to enormous 2026 capital spending, $195 billion to $205 billion for Alphabet and $130 billion to $145 billion for Meta, but they spend it on different bets. Google designs its own TPU chips, keeps Gemini closed while publishing smaller open Gemma models, and sells compute to others. Meta buys GPUs, is building its own inference chips, and has moved from open Llama to closed Muse models.

This post sets the two strategies side by side on three questions: where the chips come from, whether the models are open or closed, and who pays for the compute. It is a comparison, not a full profile of either company. For the detail on each, see the deep dives on Google DeepMind and Gemini 4 Argon and on Meta's shift from Llama to Muse.

Every figure carries its source and date. Where a point comes from a news report or a commentary piece rather than the company, the text says so. Nothing here is investment advice.

Key numbers at a glance

• Alphabet 2026 capex guidance: $195 billion to $205 billion, raised from $180 billion to $190 billion. Source: Alphabet call, 22 July 2026, via MarketBeat summary.
• Alphabet capex, April to June 2026: $44.924 billion; free cash flow negative $5.855 billion. Source: Alphabet earnings release, 22 July 2026.
• Google Cloud revenue, same quarter: $24.768 billion, up 82 percent, with backlog of $514 billion. Source: release and call summary.
• Anthropic on Google TPUs: about 3.5 gigawatts of capacity from 2027. Source: Dataconomy report on a Broadcom filing, 7 April 2026.
• Gemini 4 Argon: broadly released 13 October 2026 at an introductory $2 per million input tokens and $10 per million output tokens. Source: Google blog, 13 October 2026.
• Meta 2026 capex guidance: $130 billion to $145 billion. Source: Meta results, 29 July 2026.
• Meta free cash flow, April to June 2026: $784 million. Source: Meta results release, 29 July 2026.
• Meta and AMD: up to 6 gigawatts of AMD Instinct GPUs over several years, first gigawatt shipping from the second half of 2026. Source: AMD release, 24 February 2026.
• Meta MTIA chips: four generations announced, MTIA 300 already running ranking and recommendation work. Source: Reuters report, 11 March 2026.
• Muse Spark: released 8 April 2026, described as closed source. Source: TheNextWeb.
• Muse Wonder and Muse Wonder Full: announced 12 October 2026, offered through the Meta Model API. Source: Meta newsroom.

What "closed" and "open" mean

A model is a very large file of numbers, called weights, produced by training. A closed model keeps those weights private. Customers use it through the company's interface, usually paying per token, where a token is a small piece of text, roughly a short word or part of one. An open weights model publishes the weights, so anyone can download and run the model on their own computers, inspect it and adapt it, within whatever licence applies.

"Open weights" is not the same as "open source" in the older software sense. Training data and training code are often not shared, and licences can restrict use. When a company says "open", the useful question is what exactly is released and under which terms.

Chips: Google designs its own, Meta buys and builds

Google. Google designs its own AI chips, called tensor processing units or TPUs. The current generation and the newly announced eighth generation are covered in the TPU 8t and 8i explainer. Two developments in 2026 changed what TPUs mean for Google's business. First, in the April to June quarter, Alphabet recognised revenue for the first time from selling TPU systems for delivery into customers' own data centres, per the MarketBeat call summary. The release says Google Cloud's product revenue comes mainly from these system sales but gives no separate figure. Second, Anthropic became a very large TPU customer: a Broadcom filing reported by Dataconomy says Anthropic gets about 3.5 gigawatts of Google TPU capacity from 2027, with use tied to Anthropic's continued commercial growth. A gigawatt is a billion watts of electrical load, roughly the output of a large power station.

Owning the chip design does not make Google self-sufficient. Storage Newsletter, reporting on a SpaceX stock offering document in June 2026, says Google signed a cloud services agreement on 5 June 2026 to rent roughly 110,000 Nvidia GPUs from SpaceX's AI business at $920 million a month from October 2026 to June 2029, about $30 billion in total. That figure comes from a secondary report of the offering document, not from Alphabet. It fits what Alphabet's finance chief said on the July call: "We're still in a supply-constrained environment," with third party capacity rented in the third quarter as a bridge.

Meta. Meta's path is the reverse. It has bought most of its AI compute from chip makers and is now adding its own designs. On 24 February 2026, AMD announced a multi-year agreement for Meta to deploy up to 6 gigawatts of AMD Instinct GPUs, starting with a custom chip based on the MI450 design, with the first gigawatt shipping from the second half of 2026. AMD gave Meta a warrant for up to 160 million AMD shares that vests as shipments reach set levels and AMD's share price passes set thresholds. In the release, Mark Zuckerberg called it "an important step for Meta as we diversify our compute."

Meta also designs chips. Reuters reported on 11 March 2026 that Meta had set out four generations of its Meta Training and Inference Accelerator, or MTIA. MTIA 300 already powers ranking and recommendation, the systems that choose which posts and ads people see. MTIA 450 and 500 are aimed at inference, the work of running a trained model to answer a request, and Reuters says the remaining chips are due in 2026 and 2027, with Broadcom helping on design and TSMC making them. StorageReview reported on 13 March 2026 that Meta has "hundreds of thousands of MTIA devices in production." Reuters adds that Meta signed large deals with Nvidia and AMD in February, so its own chips complement purchased ones rather than replace them.

The difference in one line. Google sells its chips to outsiders as a product. Meta's chips are for its own data centres, and it remains a large buyer of other companies' GPUs. The wider chip market is covered in the AI chip market share guide.

Models: Google mixes closed and open, Meta moved to closed

Google. Google's frontier models, the Gemini family, are offered as hosted services. Gemini 4 Argon, released broadly on 13 October 2026, is available through the Gemini API, Google AI Studio, the Gemini app and AI Mode in Search, per Google's launch post. The post describes access only through Google's own services and does not mention downloadable weights. Introductory pricing is $2 per million input tokens and $10 per million output tokens until 30 November 2026, then $4 and $20. Google also publishes a separate family of open models called Gemma. The same launch post describes Argon being used to speed up inference for Gemma 4 E2B by 2.85 times on an Nvidia A100 GPU running the open source vLLM serving software, with the changes contributed upstream. So Google runs a two track model policy: closed at the top, open for smaller models.

Meta. Meta took the opposite route. In July 2024 Zuckerberg wrote that "open source is necessary for a positive AI future," as quoted by the commentary newsletter The Median, and Meta released a run of open weights Llama models. On 8 April 2026 Meta Superintelligence Labs released Muse Spark. TheNextWeb reports it as closed, with a private API preview, and says Meta hopes to release future versions under an open licence. Meta's own launch post says Muse Spark is available at meta.ai and in the Meta AI app but does not address the weights. In October Meta added Muse Wonder and Muse Wonder Full, offered through the Meta Model API and Meta's Muse Code product; Meta's announcement of 12 October 2026 does not mention open weights, pricing or Llama.

Meta also claims its rebuilt training stack reaches comparable capability with "over an order of magnitude less compute" than its earlier Llama 4 Maverick model. That is Meta's own claim, without independent verification.

Why the change. The Median attributes Meta's shift to the cost of capital: with spending this large, Meta needed a more direct route to a return. That is an interpretation, not a Meta statement. Meta has said only that it hopes to open future versions, with no dates. For other labs' choices, including Mistral, which publishes open weight models alongside closed ones, see the companion post on how frontier labs split AI spending.

Money: who pays for the compute

Google has three revenue streams that the same infrastructure can serve: its own products such as Search and the Gemini app, cloud capacity and models sold to companies, and TPU systems sold to other labs. In the April to June quarter Google Cloud earned $8.814 billion of operating income on $24.768 billion of revenue, and management said it expects just over half of the $514 billion backlog to become revenue within 24 months, per MarketBeat. Even so, capex of $44.924 billion exceeded operating cash flow of $39.069 billion, leaving free cash flow negative $5.855 billion for the quarter, per Alphabet's release. AI Weekly, citing Quartz, reports that Alphabet raised about $49.6 billion from a June share sale and $20.3 billion from senior notes during the quarter; that is a press report, not a figure in the release. The cloud AI revenue and backlog post sets Google Cloud against AWS, Azure and Oracle.

Meta sells no cloud service. Its revenue was $60.8 billion in the April to June quarter, with advertising at $59.4 billion, according to its results release of 29 July 2026. Its AI spending has to pay back through better ad ranking, assistants and, more recently, model access through its API. Free cash flow in the quarter was $784 million against capex of $31.08 billion. Meta ended June with $90.26 billion of cash and securities and $83.66 billion of long term debt.

To stretch its balance sheet, Meta uses partners. Storage Newsletter's summary of the results mentions a $21 billion agreement with CoreWeave for inference capacity. Meta's Hyperion campus in Louisiana sits in a joint venture with Blue Owl, and Meta described a venture with BlackRock for a 1 gigawatt site in El Paso, Texas, per MarketBeat; both are explained in the capex scoreboard. On the July call, Zuckerberg said Meta has had offers to buy its compute at a "meaningful premium" to what it paid, and argued that selling intelligence on that compute could earn higher margins.

Side by side

• Chips: Google designs TPUs and sells TPU systems to others; it also rents Nvidia GPUs when short, per a report on the SpaceX filing. Meta buys AMD and Nvidia GPUs at gigawatt scale and is adding its own MTIA inference chips.
• Frontier models: both closed. Gemini 4 Argon is hosted only; Muse Spark and Muse Wonder are offered through Meta's apps and API, with press reports describing Muse Spark as closed.
• Smaller open models: Google publishes Gemma. Meta's open Llama line has no announced successor with open weights.
• Who pays: Google's cloud and chip customers pay for part of its build out. Meta funds it from advertising cash flow, debt and partner ventures.
• Constraint: both describe more demand than supply. Alphabet said so directly; Meta's GPU deals and partner capacity point the same way.

What this means if you build on these models

• Closed models are easier to start with and carry less operational work, but the provider controls pricing, availability and retirement dates. Argon's price is set to double after its introductory period, which is a reminder to budget on the standard rate.
• Open weights models give control over where data goes and let a team pin a version, but the team carries the hosting, the security and the updates.
• A provider's strategy can change. Meta moved from open to closed within two years. If a product depends on one model, keep a written plan for switching and test alternatives on real tasks.
• Supply is limited. Capacity limits at the provider can show up as rate limits or waiting lists for customers.

What is firm and what is uncertain

Firm, from company filings, releases and posts: Alphabet's capex, cash flow and Google Cloud figures; Gemini 4 Argon's release date, channels and pricing; Meta's capex guidance, cash, debt and quarterly results; the AMD agreement terms; Muse Wonder's announcement and API availability.

Reported but not confirmed by the companies: the closed weights status of Muse Spark, Meta's plans for future open releases, the MTIA roadmap timing (Reuters and StorageReview), Anthropic's 3.5 gigawatt TPU terms (Dataconomy on a Broadcom filing), Google's GPU rental from SpaceX (Storage Newsletter on an offering document) and Alphabet's June financing (AI Weekly citing Quartz).

Interpretation: the reasons Meta changed course, and whether MTIA reduces Meta's need for purchased GPUs over time.

What to watch next

• Alphabet's and Meta's third quarter results, due in the coming weeks, especially capex guidance, free cash flow and any first view of 2027.
• Whether Google reports TPU system sales as a separate line.
• Whether Meta publishes open weights for any Muse model, and under what licence.
• Delivery of the first gigawatt of AMD GPUs to Meta, scheduled from the second half of 2026, and MTIA 450 deployment, which StorageReview reports for early 2027.
• Argon's move to standard pricing after 30 November 2026.

Quick check

If you build on any of these models, Safeguard offers SBOM and dependency scanning for the code around them, with a free scan at https://gold.safeguard.sh/scan.

Sources

• Alphabet, second quarter 2026 earnings release (SEC Exhibit 99.1), 22 July 2026: https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm
• MarketBeat, Alphabet Q2 2026 earnings call highlights: https://www.marketbeat.com/instant-alerts/alphabet-q2-earnings-call-highlights-2026-07-22/
• AI Weekly, Alphabet Q2 summary: https://aiweekly.co/alerts/alphabet-q2-cloud-jumps-82-to-248b-capex-hits-449b
• Google, "Gemini 4 Argon: Built to solve your hardest problems," 13 October 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
• Dataconomy, Anthropic, Google and Broadcom TPU deal, 7 April 2026: https://dataconomy.com/2026/04/07/anthropic-google-broadcom-strike-3-5gw-tpu-deal/
• Storage Newsletter, SpaceX IPO coverage, 11 June 2026: https://www.storagenewsletter.com/2026/06/11/spacex-ipo-is-coming-very-soon-now/
• AMD, "AMD and Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs," 24 February 2026: https://ir.amd.com/news-events/press-releases/detail/1279/
• Reuters via TimesLIVE, "Meta unveils plans for batch of in-house AI chips," 11 March 2026: https://www.timeslive.co.za/news/world/2026-03-11-meta-unveils-plans-for-batch-of-in-house-ai-chips/
• StorageReview, Meta's MTIA roadmap, 13 March 2026: https://www.storagereview.com/news/metas-mtia-roadmap-four-chip-generations-in-two-years-put-genai-inference-first
• Meta, "Meta Reports Second Quarter 2026 Results," 29 July 2026: https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx
• Storage Newsletter, Meta fiscal Q2 2026 results: https://www.storagenewsletter.com/2026/07/30/meta-platforms-fiscal-2q26-financial-results/
• MarketBeat, Meta Q2 2026 earnings call highlights: https://www.marketbeat.com/instant-alerts/meta-platforms-q2-earnings-call-highlights-2026-07-29/
• TheNextWeb, Meta Muse Spark launch, April 2026: https://thenextweb.com/news/meta-muse-spark-msl-first-model
• Meta AI, introducing Muse Spark: https://ai.meta.com/blog/introducing-muse-spark-msl/
• Meta, "Introducing Muse Wonder and Muse Wonder Full," 12 October 2026: https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
• The Median, "Meta Abandons the Open-Source Strategy?": https://dcthemedian.substack.com/p/meta-abandons-the-open-source-strategy

Last reviewed: 19 October 2026.