Swapping Meta For Google

The rationale behind it

By Ze Yao

Recently, I decided to sell 70% of my shares in Meta and instead reinvest it back into Google. My initial rationale for trimming Google was due to several reasons that can be found in the Substack writeup titled “Logging My Rationale For Trimming Alphabet.”

In that article, I stated my worries of Google being unable to increase their advertisement prices to cover the increasing costs of search queries via the LLM, which was approximated to be 10X the cost of delivering the traditional search query. However, recent updates at the investors’ presentation, as well as Alphabet’s Q1 earnings call, has alleviated my worries, as Alphabet has stated that their cost of meeting the search queries has actually dropped by over 30% thanks to their increased utilisation of their TPUs. Monetisation of their ads shown on LLM are as effective as their original search queries and asking around my fellow SMMA owners and my ex-clients also fits the narrative that google ads have been working as well-if not better than before as Google ROAS still remains higher then Meta and the gap is widening.

Furthermore, Google’s ability to be the lowest-cost provider of computing will likely be entrenched, given that the company has had a tremendous headstart in being able to design their own TPUs, which puts the company at least a decade ahead of their peers. Google can be considered the grandfather of LLMs, given that it was their technology that enabled the development of LLMs. In a paper titled “Attention Is All You Need,” Google introduced the transformer architecture to the world that enabled developers to move away from sequential processing (like RNNs and LSTMs) to a fully parallelizable approach, which is the foundational breakthrough enabling modern Large Language Models.

If Meta were to get into the cloud computing industry due to oversupply and crush margins in the industry, google will likely be able to tide through given that it is highly vertically integrated.

The core reason why Google will be able to be the lowest-cost producer in the compute industry is due to two simple reasons:

Hardware Specialization and Memory Optimization

The core of Google’s cost-efficiency lies in the evolution of their Tensor Processing Units, such as the specialized TPU 8i designed specifically for inference. Unlike general-purpose GPUs, these chips are purpose-built to handle the Transformer architecture by maximizing Matrix Multiplication Unit utilization. To tackle the memory bottleneck that plagues LLM inference, where the system is often stalled waiting for model weights to be fetched from high-bandwidth memory, Google has dramatically increased on-chip Static RAM. By keeping a larger portion of the Key-Value cache in this fast, on-chip memory, the TPU reduces the frequency of trips to slower high-bandwidth memory, significantly lowering latency and power consumption per token. Furthermore, architectural advancements like the Boardfly networking topology have flattened the communication structure between chips, reducing the latency overhead of multi-chip coordination in large model serving.

Software Stack and Algorithmic Efficiency

Beyond the hardware, Google leverages a tightly integrated software stack, including the XLA compiler, which generates highly optimized machine code specifically for the TPU architecture. This allows for advanced techniques like paged attention, which dynamically manages memory to prevent wastage, and speculative decoding, which uses a smaller, faster draft model to predict tokens that the larger target model then verifies in parallel. By optimizing the full stack from the JAX-based modeling libraries to the physical data center fabric, Google can maintain higher productive compute time and achieve better performance-per-dollar. These efficiencies allow them to pack more inference requests into the same infrastructure, fundamentally lowering the cost to serve each individual token while simultaneously improving response times for the end user.

Google’s competitive advantage stems from owning the full stack that is vertically integrated, and currently, no other company has a fully integrated stack (other than possibly Amazon). Their custom Inter-Chip Interconnect and optical circuit switching allow them to scale thousands of chips into TPU Pods that function as a single, massive computer. This level of interconnect efficiency is difficult for standard off-the-shelf ASICs or GPU clusters to match without significant specialized networking overhead. This cost advantage can be clearly seen in controlled testing for specific LLM inference tasks, where Google has reported significant gains, often cited as 2x to 4x better performance-per-dollar compared to general-purpose cloud inference offerings.

While technical proficiency is essential for navigating the current landscape, focusing exclusively on Google’s specific hardware or software architecture may be a short-sighted strategy. Because AI hardware cycles are incredibly rapid, rendering today’s state-of-the-art accelerators obsolete within a few years, investing too much cognitive capital into the mechanics of current TPUs offers diminishing returns and provides little ability to understand whether Google will succeed in the future. Instead, I believe that a more robust approach is to evaluate the underlying organizational culture. Based on my previous posts, I am sure that readers are aware of my admiration for Sundar and his team, and how they have been able to spot opportunities decades before others caught up to them and invested heavily in those ventures. By serving as an early, pivotal investor in visionaries like SpaceX, Anthropic, and DeepMind, the company has consistently placed capital behind the future’s most critical infrastructure, whether in aerospace, foundational AI safety, or general intelligence. This pattern of high-conviction and successful investing in different areas clearly shows that Google’s team has a knack for capital allocation and a disciplined approach to spend their bets evenly, as compared to Zuck’s all-in approach to AI, currently without a track record of good capital allocation (we all remember the Metaverse).

Google also has a tremendous culture that has benefited them incredibly. In 2013, Demis Hassabis and his co-founders were looking for a partner with the computational scale necessary to achieve their ultimate mission, which is to build Artificial General Intelligence. Both Google and Facebook (now Meta) emerged as major suitors.

Mark Zuckerberg was eager to acquire DeepMind, viewing it as a powerful asset for his growing empire. Reports indicate that Meta actually offered a higher financial valuation for the company, hoping to lure the founders with enormous signing bonuses and immediate capital. (Sounds familiar? History does not repeat, but it sure does rhyme.) During negotiations, Demis Hassabis famously tested the two leadership teams. When he met with Zuckerberg, he steered the conversation toward the long-term potential of AGI. However, he noted that Zuckerberg seemed equally excited about a wide range of disparate technologies and treated AGI as just one item on a list which included VR and 3D printing, whereas Larry Page demonstrated a deeper understanding of the foundational importance of AI and granted them the autonomy to pursue high-risk, long-cycle scientific research (which later culminated in AlphaFold, that enabled Demis to be awarded the Nobel Prize), whereas Zuckerberg refused to fund technology that could not be commercialised. Despite the lower financial offer from Google, Hassabis chose them and prioritized a partner whose vision for the scale and impact of AI matched his own. Looking at Zuck’s recent acquisition of “talent” via huge signing bonuses, as well as large and costly acquisitions of nascent AI founders that could be lured away by another large paycheck, shows how important culture is in an organisation. Google’s enduring competitive advantage lies in its cultural commitment to a purpose that transcends short-term economic gain. The top-tier researchers who are currently pushing the boundaries of AI and fundamental science are rarely motivated solely by financial incentives and are instead driven by the pursuit of discovery and the ambition led by a visionary leader like Demis Hassabis. This mission-centric ethos defines the core of Google’s workforce, setting a precedent that is ultimately guided by leadership whose life’s work is rooted in a vision for societal progress rather than mere capital accumulation. Demis will undoubtedly be one of the key players that will lead the future of AI.

In the midst of a massive technological paradigm shift, it is a company’s ability to foster radical innovation, maintain high-velocity engineering, and pivot its internal infrastructure that dictates long-term success. Technology will inevitably be commoditized or eclipsed by the next innovation, but a culture designed to thrive in uncertainty and adapt to shifting requirements remains the only sustainable competitive advantage, and that is what Google clearly has.

The other reason for increasing my position in Alphabet is the company’s strategic exposure to industry-leading private enterprises, specifically SpaceX and Anthropic. While I am generally hesitant to invest directly in private companies due to their elevated valuations, I am comfortable gaining exposure to these assets through an exceptionally well-run entity like Alphabet.

Meta conversely lacks these strategic stakes. It is worth considering the immense value creation that might have occurred had Mark Zuckerberg allocated capital toward such high-growth external ventures rather than focusing exclusively on the internal development of the Metaverse.

As of early 2026, Alphabet maintains an ownership stake of approximately 6.1% in SpaceX, following dilution from its merger with xAI. With SpaceX targeting an IPO valuation between $1.75 trillion and $2 trillion, Alphabet’s 6.1% stake is estimated to be worth roughly $106.75 billion. Furthermore, Alphabet holds an estimated 14% ownership stake in Anthropic. Following Anthropic’s recent Series H funding round, which valued the company at $965 billion, Alphabet’s 14% stake is valued at approximately $140 billion. Combined, these two positions represent roughly $246.75 billion of Alphabet’s market capitalization.

When factoring in Waymo, which was recently valued at $126 billion, alongside Alphabet’s interests in Stripe and various other private ventures/subsidaries (most crucially Demis’s Isomorphic Labs) that have yet to go public, the total value of the company’s non-core stakes could be worth as much as $450 billion. This valuation still excludes the upside of Google’s quantum computing research, the clear leader in the field. By maintaining such a diverse portfolio of foundational technology, Alphabet offers a level of strategic optionality that distinguishes it from Meta and provides a unique right-tail upside potential in nascent industries that no other company is able to provide, while its core business remains highly profitable and firing at all cylinders.

Alphabet represents a collection of significant call options on industries that remain largely undefined given that both the nature of these sectors and their long term economics are currently beyond our foresight. As investors, we do not need to possess the expertise to navigate these emerging fields, and we can instead rely on management to allocate capital effectively given their superior circle of competence. This strategy allows shareholders to benefit from the massive cash flows of the core advertising business, which provides the necessary fuel for Alphabet to fund potential breakthroughs in technology and beyond. Provided that leadership maintains a disciplined approach to capital expenditure and continues to prune underperforming projects, the company remains a powerful vehicle for capturing value in unknown future markets. By investing in Alphabet, I am able to harness the judgment of Sundar and team, who are uniquely positioned to identify and invest in significant opportunities.