This strategy, where specialised firms cooperate and compete across the AI stack, was lauded as a nimble, genius business model. But let us be clear: this model was never about superior agility; it was about leveraging market dominance, and that dependence has now revealed its crippling limits.

Look at the evidence: on October 28, 2025, Microsoft was forced to loosen its grip on OpenAI, granting the lab “boundless promiscuity” in choosing cloud partners, waiving Microsoft’s prior right of first refusal—no permission required.

This shocking structural concession came despite a monumental, restructured deal that guaranteed Microsoft a massive 27% ownership stake in OpenAI’s for-profit entity, valued at approximately USD 135 billion, and entitled the tech giant to 20% of the startup’s revenue until 2032, or until an expert panel verifies the achievement of Artificial General Intelligence (AGI).

The hypocrisy is stunning. Microsoft secured a USD 250 billion commitment from OpenAI for Azure cloud services over the next few years, yet OpenAI is simultaneously forced to make massive infrastructure deals with rivals.

It is contracting with Oracle for data centre services, joining a USD 500 billion data centre project called “Stargate,” and securing compute resources from AMD for six gigawatts of GPUs. The truth is that the intense, multi-billion-dollar cost of running frontier AI models—costs OpenAI is desperately trying to alleviate—is simply too high for any single horizontal partner to bear, even one receiving a quarter of a trillion dollars.

This crippling reliance on external suppliers for core compute is not a triumphant strategy; it is a debilitating structural weakness, forcing the supposed AI king to scramble for capacity across rival clouds.

Google’s Silent Triumph
In stark contrast to Microsoft’s frantic horizontal scrambling, Google, the one company universally mocked as the AI laggard, went all in on deep vertical integration—a strategy once dismissed as slow and burdened by bureaucratic inertia.

Google is the only major technology company that has committed entirely to this end-to-end approach, designing its own Tensor Processing Units (TPUs) in-house, training frontier models through Google DeepMind, and deploying them across its massive ecosystem, including Search and YouTube.

While corporate inertia did cause a delay in releasing their chatbot, leading to initial embarrassment, that delay was architectural patience, allowing them to refine a system that is now economically devastating to every competitor.

Google’s vertical control over its hardware stack has provided a structural, defensible advantage, a true moat that no amount of market spending can easily breach. The proprietary TPU is not just faster; it is a superior financial weapon.

TPUs are reported to be 4 to 10 times more cost-effective than conventional GPUs in large-scale language model training scenarios, offering 1.2x to 1.7x better performance per dollar compared to NVIDIA A100 GPUs. Furthermore, due to optimised deep learning architecture and reduced cooling requirements, TPUs consume 30% to 50% less power, leading to vast cost savings for Google and achieving up to 2.7x better performance per dollar in some instances.

This architectural genius means the cost per AI query for Google is not five times that of traditional search, as the doom-mongers originally estimated, but only twice. This control over unit economics is everything—it means integrating AI into Search dilutes Google’s gross margin only marginally, from a robust 90% to a still-highly competitive 86%. This is the economic foundation of sustained power, a foundation Microsoft simply does not possess.

Why Verticality Always Wins
The market has now, finally, recognised this strategic gulf, turning its cold shoulder toward Alphabet into a warm embrace. This reversal resulted in a USD 1 trillion gain in market value over just four months, supported by strong financial performance that proves AI is additive, not cannibalistic, with Google Cloud sales growing at an annual rate of 35%, driven explicitly by generative AI solutions and infrastructure.

The most damning indictment of the horizontal model, however, comes directly from Google’s competitors. Anthropic, a prominent AI lab, has now become a major, deeply committed customer of Google’s proprietary infrastructure, announcing a massive expansion to utilise up to one million TPUs from Google Cloud, a deal worth tens of billions of dollars.

This move was driven explicitly by the superior “price-performance and efficiency” of the TPUs, confirming that Google’s vertically integrated approach has created an economic necessity that rivals cannot ignore, compelling them to become major customers for the infrastructure they were supposed to be bypassing.

Recognising this fundamental failure, Microsoft is now frantically pivoting to mirror Google’s strategy, establishing its own in-house chip-design studio and AI lab in a belated effort to control its destiny.

But attempting to copy a mature, decade-long architectural advantage is costly, slow, and strategically embarrassing. Microsoft’s second-generation Maia chip is already delayed, and its model-building efforts are considered “inchoate.”

The war for the AI stack is, for all intents and purposes, fundamentally decided, and the vertical approach, powered by proprietary silicon, has won—comprehensively—through superior unit economics.

The post IF Insights: Google vs Microsoft, the battle for infrastructure power appeared first on International Finance.

The Series B round was led by its existing investors, including Walden Catalyst Ventures, 83North, Koch Disruptive Technologies, Pitango First, and Viola Ventures, as well as strategic investors, including Lip-Bu Tan, CEO of Intel and managing partner at Walden Catalyst Ventures, and Eyal Waldman, co-founder and former CEO of Mellanox Technologies.

The APU architecture focuses on addressing the specific bottlenecks of analytics at the computing level, unlike graphics processing units (GPUs) developed by the existing semiconductor sector players, which were initially designed for graphics and later modified for AI and data-related tasks. In today’s episode of the “Start-up of the Week,” International Finance will talk about Speedata and its game-changing APU.

Transforming The Data Centre Landscape

Speedata claims its purpose-built APU to be the first of its kind, specifically designed for accelerating database analytics and AI workloads. The solution also offers unparalleled scalability and efficiency, seamlessly integrating hardware and software into a powerful, unified platform.

“For decades, data analytics have relied on standard processing units, and more recently, companies like Nvidia have invested in pushing GPUs for analytics workloads. But these are either general-purpose processors or processors designed for other workloads, not chips built from the ground up for data analytics. Our APU is purpose-built for data processing and a single APU can replace racks of servers, delivering dramatically better performance,” said Adi Gelvan, CEO of Speedata, in an interview with TechCrunch.

The start-up was founded in 2019 by six founders (including Gelvan), some of whom were the first researchers to develop Multi-Threaded Coarse-Grained Reconfigurable Architecture (CGRA) technology.

The founders collaborated with ASIC design experts to address a fundamental problem related to the data analytics being performed by general-purpose processors. If the workloads grew too complex, they could need to tap into hundreds of servers. Gelvan and his colleagues believed that they could develop a single dedicated processor to accomplish the task faster using less energy.

Talking about Speedata’s APU, the product currently targets Apache Spark workloads (an open-source, distributed processing system used for big data workloads). As per Gelvan, in the long run, the unit will support every major data analytics platform.

“We aim at becoming the standard processor for data processing. Just as GPUs became the default for AI training, we want APUs to be the default for data analytics across every database and analytics platform,” Gelvan stated.

The start-up reportedly has several large companies testing its APU. Speedata also claims a specific case where its APU completed a pharmaceutical workload in 19 minutes, which was significantly faster than the 90 hours it took when using a non-specialised processing unit, resulting in a 280 times speed improvement.

The start-up also has achieved several milestones since its last fundraising, including finalising the design and manufacturing of its first APU in late 2024.

“We’ve moved from concept to testing on a field-programmable gate array (FPGA), and now we are proud to say we have working hardware that we are currently launching. We already have a growing pipeline of enterprise customers eagerly waiting for this technology and we’re ready to scale our go-to-market operations,” Gelvan said.

Meet The Game-Changing APU

The Speedata APU achieves its breakthrough throughput by mapping the required processing into its internal hardware pipeline. The Speedata Dash software automatically configures a data flow in silicon, where row processing is broken into hundreds of steps, each efficiently flowing to the next one at every hardware clock. Therefore, at any given time, hundreds of rows are at different stages of processing in the hardware, in parallel, resulting in a processing throughput of over a billion rows per second.

Speedata APU also ensures accelerated Parquet processing in hardware. Parquet is the leading file format for analytics. Speedata’s APU efficiently processes Parquet files as part of its hardware pipeline, from decompressing and decoding columns through columnar filters and projections to rows assembly and flattening of nested data (EXPLODE). And then comes Speedata APU’s seamless integration with Apache Spark.

“Speedata’s Dash software transparently plugs into the Spark Catalyst optimiser to automatically identify and offload compute-intensive work to the APU, delivering dramatic acceleration for Apache Spark 3.x workloads on Kubernetes, YARN and standalone cluster managers,” the start-up stated.

Speedata C200 is the industry’s first analytics accelerator card, housing the start-up’s APU chip.

“It is optimised for high-bandwidth data processing and delivers dramatic acceleration for Apache Spark workloads. With its PCIe connectivity, it is engineered for seamless integration into standard server configurations,” Speedata noted.

With flexible server deployment options, Speedata makes it easy to bring the power of the APU to data centres, with customers having the freedom to choose between two flexible procurement models. One of them is buying pre-configured from the start-up, under which the latter provides a fully integrated 2U server with two C200 analytics accelerator cards, pre-configured and ready to accelerate Apache Spark workloads out of the box.

Under the second option, Speedata offers data centre servers equipped with C200 analytics accelerator cards through trusted OEM providers such as HPE. These servers can be customised in terms of model and configuration to meet the client’s needs.

To ensure seamless integration with Apache Spark, Speedata’s Dash software transparently plugs into the Spark Catalyst optimiser to automatically identify and offload compute-intensive work to the APU. It automatically delegates to the CPU the handling of SQL (Structured Query Language) elements such as UDFs (User-Defined Function) that cannot be accelerated, while still handling the rest of the query, guaranteeing great performance.

The Workload Analyser is a standalone tool for analysing Apache Spark event log files and identifying the projected acceleration that Speedata delivers for its clients’ workloads. The latter can quickly learn which queries will benefit the most, whether a faster network will have a big impact on the business’ APU environment or not, or see a detailed per-stage analysis of the benefits or limits.

The post Startup of the Week: Speedata challenges chip giants with disruptive APU architecture appeared first on International Finance.