The $50 Billion Entry Fee: How Fiercely Are Tech Giants Competing to Break into Healthcare?

“A brand-new computing platform has emerged,AI’s ‘iPhone Moment’ Has Arrived“, accelerated computing and AI technology have become a reality,” summarized NVIDIA founder and CEO Jensen Huang, beaming with confidence, in his opening keynote at the GTC 2023 Spring Conference. At this year’s GTC, there were more than 70 sessions focused on generative AI.

This perspective is not an exaggeration. Just one day ago, Microsoft and Nuance unveiled a killer AI application that “iPhone-ifies” clinical documentation for physicians—the first healthcare application to integrate the GPT-4 model, launched less than a week after GPT-4’s release. Microsoft’s record-breaking $10 billion investment in OpenAI now appears exceptionally worthwhile, not only marking a highlight moment for Microsoft in the healthcare sector but also symbolizing the broader trend of tech giants steadily increasing their stakes in healthcare.

The day before the release of GPT-4, Google unveiled a generative AI model specifically designed for healthcare. Although it arrived later than the wildly popular ChatGPT, generative AI has substantial room for application in the medical field. This marathon is not about who starts first, but about who can endure until the end.

Tech giants already in the market are continuing to deepen their engagement, while those yet to enter are seizing every potential opportunity to break in. In response, semiconductor giant AMD has also emphatically said “Yes.” After all, for the first time in decades, it has stamped its arrow logo onto the healthcare sector.

What Will the Tech Giants’ Intensified Competition in Healthcare Bring Us? It’s Worth Anticipating.

Microsoft’s foray into the healthcare sector actually began quite early; as far back as 1999, it invested in the medical information website WebMD. In the subsequent years, Microsoft also acquired several healthcare-related companies. However, at that time, Microsoft did not treat healthcare as a core business, nor did it integrate these operations.

In 2017, Microsoft launched Healthcare NExT, marking its systematic entry into the healthcare sector. This initiative focused on the then-emerging field of medical AI, leveraging Microsoft’s strengths in artificial intelligence and cloud computing to drive healthcare innovation. Its primary objectives included alleviating physicians’ data entry burdens, triaging patients, and monitoring out-of-hospital patient care.

In the following years, Microsoft largely continued along this trajectory in the healthcare sector. In 2020, it launched its first cloud computing solution specifically designed for the healthcare industry—Microsoft Cloud for Healthcare.

However, to be frank, Microsoft’s reputation in the healthcare sector is not as prominent as that of Nuance, which it later acquired. Nuance was once a leader in the field of voice AI, serving not only as the developer of Apple’s Siri voice engine but also commanding over 60% of the global intelligent voice market share. After facing challenges from tech giants, Nuance shifted its business focus to the healthcare industry and established significant competitive barriers through years of dedicated cultivation.

In April 2021, Microsoft announced its $19.7 billion acquisition of Nuance. This transaction marked the third-largest acquisition in Microsoft’s history since its founding, significantly strengthening its influence in the healthcare vertical.

Nuance’s services have been embedded into Microsoft’s healthcare cloud, providing physicians with speech recognition and transcription capabilities. The voice AI intelligently recognizes doctor-patient conversations, performs contextual analysis, and then inputs the data into electronic health records (EHRs) to automatically generate clinical documentation, thereby improving the efficiency of physicians’ diagnostic workflows.

Nuance’s DAX Can Generate Clinical Documentation via Voice (Image from Nuance’s Official Website)

The integration of GPT models will significantly reduce the time required to generate clinical documentation.Generally, generating clinical documentation with DAX without the GPT-4 model takes approximately four hours. Leveraging the powerful generative large language model and reasoning capabilities of GPT-4, DAX Express reduces this process to just a few seconds.。

This significantly enhances physicians' user experience, reduces their administrative burden, enables real-time clinical documentation, and improves efficiency.

For Microsoft, being the first to integrate GPT-4 will significantly boost its influence in the healthcare sector—now, orders of magnitude more people are aware that Microsoft is also deeply committed to healthcare.

This preemptive move is not difficult to understand. Microsoft has continuously invested in OpenAI over the past few years, adding a reported $10 billion in investment in late January; it also provides Azure cloud services and even built a supercomputer specifically for OpenAI to train GPT-3 on the Azure public cloud. There are only a handful of commercial entities worldwide capable of providing the massive computing power required for such training.

Compared to GPT-3.5, the model underlying ChatGPT, GPT-4 demonstrates substantial improvements in both problem-solving capabilities and accuracy. GPT-4 achieved a high score of 163 out of 180 on the U.S. SAT, surpassing 88% of test-takers. Furthermore, it outperformed 90% of candidates on the bar exam, whereas GPT-3.5 performed at a level comparable to the bottom 10%.

However, the public has also observed certain issues with generative AI during their use of ChatGPT. In particular, its tendency to produce “plausible-sounding but fabricated information” can lead to significant cognitive biases and misinformation. After all, the essence of this technology lies in continuously generating data by maximizing probabilities, rather than employing algorithmic models that fully rely on logical reasoning to answer questions. Furthermore, given the certain degree of autonomy exhibited by generative AI, data security and privacy leaks are considered major concerns.

In any case, amid the race among tech giants to ramp up their investments in healthcare, Microsoft’s accelerated pace has given it a half-length lead.

Just one day before Microsoft entered the healthcare sector with GPT, Google held its annual “The Check Up with Google Health” event to showcase its latest advancements in healthcare, with medical AI being a key focus.

Before ChatGPT rose to prominence, Google was widely regarded as the leader in the field of AI. The buzz surrounding DeepMind and AlphaFold at that time was no less intense than the current excitement around OpenAI and ChatGPT. Within just a few weeks of ChatGPT’s launch last year, Google and DeepMind jointly released Med-PaLM.

This is a large language model specifically designed for the healthcare sector. Over the past few years, Google has been conducting research on medical large language models to address the professionalism and specificity required in the healthcare field.

The Med-PaLM model has achieved remarkable results. It is the first AI to attain a “passing score” (>60%) on U.S. medical licensing-style questions, not only accurately answering multiple-choice and open-ended questions but also providing rationales for its answers and self-evaluating its responses.

At the event, Google unveiled Med-PaLM 2. According to Google, the new model demonstrated an 18% performance improvement over its predecessor, significantly outperforming other comparable AI models (noting that GPT-4 had not yet been released at the time).In medical examinations, Med-PaLM 2 has largely approached the level of "expert" physicians, achieving a score of 85%.。

However, Google has also acknowledged that while these advances are encouraging, there remains a significant gap before deployment in real-world settings. Clinicians and non-clinicians from diverse backgrounds and countries evaluated the model against 14 criteria, including scientific validity, accuracy, medical consensus, reasoning, bias, and potential harm. The Google team found that Med-PaLM 2 still falls short in answering medical questions and meeting Google’s standards for product excellence. Moving forward, Google will continue to collaborate with researchers and the global medical community, aiming to ensure this technology truly helps improve healthcare delivery.

Although Med-PaLM 2 is not yet ready for practical application, Google has made significant progress in AI-powered medical imaging. Last October,Google Health Reaches Agreement with iCAD to Integrate Google’s Breast Imaging AI Technology into Its Breast Imaging Products, Marking Google’s First Deployment of Its Breast Imaging AI Model in Clinical Practice。

At the event, Google also announced that it would further strengthen the development of its AI models for ultrasound, targeting applications such as maternal care and early breast cancer screening, thereby helping to address the widespread shortage of sonographers and the inconsistent quality of ultrasound examinations globally.

Meanwhile, Google is also collaborating with Mayo Clinic to leverage AI for radiotherapy planning. Furthermore, Google will promote AI-assisted early screening for tuberculosis in Africa, where statistics show that over 25% of global tuberculosis-related deaths occur.

Although Google’s progress in healthcare appears bustling, it pales in comparison to Microsoft’s impressive performance during the same period, underscoring the adage that “no comparison means no pain.”

In fact, Google’s forays into the healthcare sector in recent years have not gone as planned. In 2021, the tech giant experienced a major upheaval in its operations—Google announced a shift in its healthcare strategy, dissolving the once-independent Google Health and integrating its personnel into other existing business lines.

Commercialization has long been a challenge for Google in the healthcare sector. In Google’s annual reports, the “Other Bets” segment refers to emerging businesses at various stages of development, including healthcare and internet services. Revenue from this business unit has seen little growth over the years, remaining steady at approximately $600–700 million; meanwhile, its losses have increased year by year, reaching as much as $5.3 billion in 2021.

Organizational structure leading to business overlap may also be a reason for Google's mediocre performance in the healthcare sector in recent years. Since Google restructured into Alphabet in 2015, its various healthcare product lines have gradually coalesced into independent business units. By early 2021, Google’s healthcare operations were roughly divided into three units: Google Health, Verily, and Calico. Long-standing overlaps among these units’ businesses have further exacerbated already challenging commercialization efforts.

Under the new architecture, whether Google can find its footing in the healthcare sector remains to be further validated. However, its case also illustrates the unique nature of the healthcare industry—even invincible tech giants struggle to navigate it effectively.

GPT, which has recently gone viral across the internet, is a type of large language model. The term “large” is evident from the scale changes in GPT models in recent years. The initial version of GPT, released in 2018, had only 117 million parameters and was pre-trained on just 5 GB of data. In contrast, GPT-3’s parameter count surged to 175 billion, with its pre-training dataset reaching 45 TB (1 TB = 1,024 GB).

GPT-4, which has already been unveiled, has not disclosed its corresponding metrics. However, given OpenAI’s typical approach and the model’s significantly superior performance compared to its predecessor, it is highly likely that its scale is substantially larger. This necessitates immense computational power for both model training and inference, entailing extremely high costs.

According to media reports, the supercomputer built by Microsoft for OpenAI’s model training features 285,000 CPU cores and more than 10,000 V100 GPUs, which were top-of-the-line at the time. Even if the V100s were replaced with the currently mainstream A100s, a scale of 3,000 units would still be required.

Furthermore, a single training run of GPT-3 cost over $4.6 million, with corresponding cloud resource expenses amounting to nearly $100 million annually. Throughout 2022, OpenAI spent more than $500 million on training GPT-3. Few companies worldwide possess the financial capacity to sustain such a “cash-burning beast” that temporarily generates no revenue. This is also a key reason why large language models of this scale failed to gain widespread adoption previously.

As generative AI gains increasing recognition, the demand for computing power is surging. According to OpenAI’s estimates, since 2012, the computational requirements for training leading global AI models have doubled approximately every three to four months, with an annual growth factor of up to tenfold.

For NVIDIA, a provider of foundational AI computing power, there could be no better news. Amid the generative AI frenzy sparked by ChatGPT,NVIDIA’s stock price has risen by nearly 80% since the beginning of the year, reaching $271.91 (as of March 23), with its market capitalization approaching $650 billion., surpassing Tesla in one fell swoop.

NVIDIA Founder and CEO Jensen Huang Says AI’s “iPhone” Moment Has Arrived (Screenshot from Live Video Stream)

At the 2023 Spring GTC Conference held recently, there were more than 70 thematic sessions related to generative AI. Meanwhile, NVIDIA swiftly launched the H100 NVL GPU, specifically designed for generative AI. Although there was not enough time to develop an entirely new architecture, linking two of the current top-tier H100 GPUs still enables a 2–3x increase in computing power.

NVIDIA and Japan’s Leading Trading House Mitsui & Co. Collaborate to Develop Tokyo-1 Supercomputer, to leverage generative AI to empower drug research and development in Japan’s pharmaceutical industry and startups. Leading Japanese pharmaceutical companies, including Astellas Pharma, Daiichi Sankyo, and Ono Pharmaceutical, have all planned to adopt this supercomputer.

Meanwhile, NVIDIA has also launched the NVIDIA AI Foundations cloud service to help customers who need to build, refine, and run custom large language models and generative AI accelerate the adoption of generative AI. This service includes NeMo for natural language processing, Picasso for image processing, and BioNeMo for the pharmaceutical industry.

Among these, BioNeMo can accelerate the most time-consuming and costly stages of the drug discovery process. It provides pre-trained models and leverages proprietary data to customize models for various stages of the drug development workflow—including generative AI tools such as ProtGPT-2, which assists researchers in designing novel proteins.

Through these models,BioNeMo can help researchers identify the correct therapeutic targets, design molecules and proteins, and predict their interactions within the human body, thereby facilitating the development of optimal drug candidates.。

Generative AI models can rapidly identify potential drug molecules and, in some cases, even design compounds or protein-based therapeutics from scratch. Trained on large datasets of small molecules, proteins, and DNA and RNA sequences, these models can predict the 3D structures of proteins and the extent to which molecules dock with target proteins.

Currently, many pharmaceutical companies are attempting to use generative AI to design new drug candidates. Amgen has utilized BioNeMo to accelerate its biologics development process, aiming to explore and develop therapeutic proteins for next-generation drugs.

Beyond drug discovery, generative AI can also enhance efficiency and optimize therapeutic outcomes in other stages of drug development, such as drug design and dose selection. By 2040, generative AI is projected to deliver $1 trillion in value to the healthcare industry.

Through years of sustained investment, NVIDIA has successfully built a massive AI foundation in the healthcare sector, fostering mutual empowerment with the medical ecosystem. On one hand, by continuously providing foundational computing power through hardware upgrades and incorporating software and services to help customers rapidly deploy applications, NVIDIA is steadily tapping into the visibly enormous “gold mine” of medical AI. On the other hand, medical AI has achieved significant advancements thanks to the support of NVIDIA’s hardware and software solutions, enabling it to play a more substantial role across a wider range of scenarios.

In 2022, another prominent name joined the ranks of tech giants venturing into the healthcare sector: AMD. As a long-standing rival of Intel and NVIDIA, AMD is also a provider of foundational AI computing power. However, it had not previously attempted to enter the healthcare field in its specific business operations.

In October 2020, AMD announced its $35 billion acquisition of Xilinx, the FPGA giant. As Xilinx was the largest FPGA manufacturer (with a market share as high as 49% in 2020), this acquisition triggered prolonged antitrust reviews across various countries and was not officially completed until 2022.As Xilinx’s stock price rose during this period, the M&A deal size reached nearly $50 billion, setting a new record for mergers and acquisitions in the semiconductor industry.。

Xilinx, with its decades-long expertise in FPGA technology, has accumulated a substantial customer base across industries such as healthcare, aerospace, automotive, and defense. This legacy is invaluable to the new AMD, securing its entry into numerous high-barrier niche markets, including healthcare.

FPGAs have long been widely used in the field of medical imaging. Major medical device manufacturers around the world are extensively utilizing FPGAs. Their applications primarily include ultrasound imaging, 3D rendering, X-ray data acquisition, and magnetic resonance imaging (MRI) interfaces. Furthermore, FPGAs can also be found in endoscopes, defibrillators, and patient monitors. In these applications, FPGAs offer greater computational power than DSPs, enabling high-speed execution of certain algorithms during the imaging process.

As a leader in the FPGA field, Xilinx has been deeply engaged in these applications for many years.Through the acquisition of Xilinx, AMD has also managed to imprint its “arrow” logo onto the medical field for the first time.。

AMD Enters the Healthcare Sector Through Acquisition (Screenshot from Event Poster)

At CES in early 2023, AMD released its first Medical Imaging Library for the Vitis 2022.2 unified software environment to deliver improved ultrasound imaging quality, as enhancing both the quality and accuracy of ultrasound imaging is becoming increasingly challenging.

AMD’s medical imaging library can significantly enhance image quality and reduce latency. By deploying ultra-high-performance ultrasound beamforming algorithms on Xilinx FPGA hardware, ultrasound systems can achieve high-quality, low-latency imaging even in challenging abdominal or cardiac scanning scenarios.

Furthermore, this medical library adopts a three-tier design, offering everything from Level 1 module construction to Level 3 complete ultrasound beamformer designs. This provides equipment manufacturers with a wealth of options, helping them shorten development cycles and accelerate time-to-market.

Frankly speaking, AMD’s current application scenarios in the healthcare sector remain relatively limited, but its future prospects are promising. This is because FPGAs, when paired with specialized algorithms, offer flexible configurability and deliver performance comparable to GPUs in specific environments, making them an alternative choice for AI companies.

For instance, in scenarios such as assisted disease diagnosis (including chronic disease screening and risk assessment), future disease risks can be predicted by learning from historical data of patients with chronic conditions. In such scenarios, the use of FPGAs can also enhance model training and inference performance.

Combined with its existing CPU and GPU product lines, AMD has significant untapped potential in the healthcare sector.

It is hardly surprising that tech giants are ramping up their investments in the healthcare sector.First, the scale of the healthcare market is truly substantial.According to statistics, the U.S. healthcare market size reached $4.09 trillion in 2021. This figure includes expenditures across multiple sectors such as healthcare facilities and services, pharmaceuticals, medical devices and products, and health insurance, but does not include research and development spending related to healthcare.

Meanwhile,Healthcare expenditures are often inelastic necessities.This is not difficult to understand—after all, it is not uncommon for people to exhaust their financial resources to seek medical treatment, but it is indeed rare for them to do so merely for fleeting pleasure. It is precisely for this reason that the consumer technology sector suffers a far greater impact than the healthcare sector during economic downturns. Unfortunately, this is exactly what is happening now: according to a World Bank research report, the global economy is currently experiencing its most severe downturn since 1970.

Nevertheless, the number of tech giants whose performance in the healthcare sector over the years has justified their investments is exceedingly rare. VCBeat believes that there are generally several reasons for the underperformance of tech giants in the healthcare field.

Leveraging their inherent attributes, tech giants typically enter the healthcare sector by extending their technological strengths and adopting innovation-driven approaches.. Once successful, such innovation is bound to establish a solid competitive advantage, as exemplified by NVIDIA’s tremendous success in AI acceleration based on GPUs. Unfortunately, the failure rate of technological innovation is extremely high, often requiring multiple attempts.

Tech Giants Find a Dramatically Transformed Landscape Upon Entering the Healthcare Industry, patient privacy protection and data collection standards may not have been major issues in the past, but they often become significant barriers in the heavily regulated healthcare sector. Previously, it was difficult to address the most complex challenges in healthcare through data-driven approaches.

Due to previous achievements,Tech giants often fall prey to the myth of their past successes, naively believing that technology can solve all problems in the healthcare sector.. Unfortunately, for patients, the issue lies more with the overall experience than with the technology itself. In fact, people tend to be conservative when it comes to healthcare matters, preferring proven, mature, and reliable solutions; consequently, the adoption and promotion of new technologies often require a considerable period of dormancy.

Unlike technology companies focused on healthcare,Hard-to-profit healthcare businesses receive only a small share of resources from tech giants, nor is there a coherent enterprise-level strategy, making it difficult to consolidate fragmented healthcare operations scattered across various departments into a unified force. There is also a lack of psychological preparedness for long-term investment in healthcare.The business unit’s strategy has to change frequently, making it unsustainable. This approach is unworkable in the healthcare sector, which demands long-term, stable investment and considerable patience.。

How to integrate one’s own strengths with the unique characteristics of healthcare to achieve a breakthrough and attain both critical acclaim and commercial success in the medical field is a question worthy of deep reflection for tech giants that have already entered or are planning to enter the healthcare sector.

We look forward to tech giants playing a more active role in the healthcare sector, thereby making greater contributions to driving the digital transformation and intelligent upgrading of the healthcare industry.