Uncertainty is escalating. From the White House to the claims made by AI company DeepSeek, predicting the future has never been more difficult. We see potential for unintended side-effects of US policy creating opportunities for China to increase its global influence; but, at the same time, how the US conducts itself over Ukraine will impact events both in Korea and Taiwan. None of this is helped by the huge boost to China’s reputation as an AI powerhouse provided by DeepSeek. However, on closer examination, we find that Beijing’s claims are substantially overstated. Against this backdrop, we see further restrictions related to AI exports and an acceleration of the bifurcation of the technology industry into 2 separate camps as China pushes harder for technology independence. Those in the middle will be forced to choose.

• Chaos and opportunity: Trump’s deliberately chaotic approach to governance has thrown up proposals in a range of policy areas which run counter to maintaining a tech lead over China.
• Tariffs & withdrawals: The imposition of tariffs on allies as well as policies like the watering down of anti-trust enforcement may help China secure an advantage over the US.
• Russia, Ukraine & Taiwan: The Trump administration’s stance on Russia/Ukraine will have an impact on events in both the Korean peninsula and the Taiwan Strait. Furthermore, changes to the US factsheet on Taiwan could indicate Washington’s preparedness to use it as a bargaining chip in a bigger game with China.
• DeepSeek: shocked the world in being able to match the performance of Western competitors but is not novel in terms of performance or intelligence.
• DeepSeek efficiencies are where the real surprises lie. DeepSeek’s claims seem to suggest that it is many orders of magnitude cheaper to both train and run inference than Western peers.
• The reality: RFM’s analysis suggests that DeepSeek’s claims are misleading and that the reality is that DeepSeek has something more like a 2x-4x advantage in training (rather than 100x) and a 0x – 10x advantage in inference depending on the task being inferenced.
• DeepSeek’s IP: is in its methods not the models it has placed in open source. This is how we think DeepSeek was able to obtain an export license for R1 from the Chinese state.
• China’s motivations: DeepSeek is operating with state approval leading to the question of China’s motivations. RFM sees three possibilities with PR and propaganda being the most likely where we think China has succeeded way beyond its own expectations.
• Build the wall 2.0. The now unfolding ideologically-driven bifurcation of the Internet will accelerate the ongoing shift from global standards to one for China, and one for the West with everyone else being forced to pick a side. Long-term growth will suffer as a consequence.

RFM Subscribers will receive their copy by email. 

Voice has long been proposed as a man-machine interface, but for decades it has failed to meet the requirements for mass market adoption. In 2017 RFM proposed three stages that AI needed to reach and with LLMs, stage 3 has finally been passed. Combine this with the low latency achieved by deploying the agent at the edge, and humans will finally be able to have meaningful conversations with machines. RFM thinks that this could have the greatest impact in the automotive sector, but smartphones, The Metaverse and many other industries are also in line to be substantially impacted.

Selling a car has never made economic sense for Apple but selling an infotainment unit does. It also gives Apple the ability to monetise digital services in the vehicle at the expense of the OEM. Apple has found a way to safely run the instrument cluster on the iPhone opening the way for a complete takeover of the digital cockpit. This will exclude OEMs from the digital vehicle leaving them with very little with which to offset revenue losses from lower vehicle shipments.

• Apple has become the largest threat. Until 2022, Apple was non-threatening to the OEMs as it was using CarPlay primarily to drive loyalty to its ecosystem. With its 2022 move into the instrument cluster, Apple signalled an intent to run the car on the phone.
• Reliability and safety. The problem is that the smartphone is not reliable enough to be considered safe to run critical vehicle systems such as the instrument cluster, presenting Apple with a problem. The solution to this problem was presented at WWDC 2024
• CarPlay 3.0. is how Apple can take control of the instrument cluster via the iPhone and still meet the reliability and safety requirements of the automotive industry. Apple splits the instrument cluster into 4 layers and renders the vital layers with Apple software and hardware on the infotainment unit while running non-critical layers from the iPhone. These are amalgamated by Apple software on the infotainment and then passed to the OEM system to add overlay items like indicators and hazard warnings.
• Software Development: is executed by the OEM using an Apple software development kit (SDK). Apple offers a lot of flexibility in terms of user experience elements but at the end of the day, the instrument cluster is clearly an Apple product with OEM customisations. This greatly reduces the OEM’s digital relevance in the vehicle and threatens to cut them out from the vehicle digital services economy entirely.
• Consumer preferences. Apple has stated that 79% of US consumers will only consider buying CarPlay-enabled vehicles but a 2023 survey from McKinsey puts the figure at closer to 40%. Regardless of which figure is most accurate, Apple has already gained a meaningful amount of market power which could be enough to convince the OEMs to integrate more Apple technology into their vehicles.
• Infotainment takeover. The logical end game for Apple is to have the OEM buy the entire infotainment unit from Apple and integrate it into their vehicles. Unlike the Apple Car, this is a viable business model that would fit well with Apple’s other businesses and give Apple the ability to monetise digital vehicle services at the expense of the vehicle maker. From an OEM perspective, this is best avoided.

RFM research subscribers will receive their copy by email. 

It turns out that AI needs a lot more than attention to create hyper-intelligent machines. RFM believes that the missing link is causality which means that the machines still have no understanding of anything that they do. This prevents the development of any form of AGI and there is no sign that the causality problem will be fixed anytime soon. However, even without causality, there are many use cases for AI-enabled by the advances in language performance and the ability to make sense of unstructured data. Unfortunately, this is not enough to meet the hype that surrounds AI and so some form of correction is needed to bring expectations and reality back into balance.

• What AI is: AI describes a range of technologies that enable machines to make decisions. They range from statistics at one end, through rules-based software, deep learning and now, generative AI. Generative AI is a subset of AI and is defined as AI that is capable of generating content such as words, images and now, video.
• What AI is not: Large Language Models (LLMs) have enabled machines to converse in natural language which has led to an extraordinary increase in anthropomorphism. This is the attribution of human characteristics to non-human objects which in this case are statistics-based algorithms. It is here where expectations for AI and reality begin to diverge.
• Causality: is by far the biggest limitation of generative AI as these models have no causal understanding of what it is that they are doing. This is what causes the machines to invent facts, make simple mistakes and remain unaware that they are doing so.
• Reasoning: this is a hot area of debate where owners of models claim they can reason, and the sceptics disagree. Reasoning is crucial because it will be the first sign of LLMs being able to understand causality. The balance of empirical evidence suggests that the machines remain unable to reason and there is no evidence to suggest that it will be solved anytime soon.
• Use cases: despite the limitations, RFM sees many use cases given that generative AI represents a large step forward in both the ability to use natural language and data characterisation, storage and retrieval. These use cases do not replace humans but make them more productive meaning that the workforce is not going to be replaced although some adaption will be needed in many industries.
• AI bubble: There is little doubt that valuations and expectations are too high. The flood of capital into generative AI has been driven by intense public excitement. The result is becoming a flood of supply of LLM services all of which are roughly equivalent in terms of performance meaning competition and falling prices. It is this that RFM expects will trigger the reset to reality. It is important to note that this reset will not be nearly as harsh as autonomous driving because generative AI has products now where autonomous driving still has nothing.

RFM research subscribers will receive their copy by email

The OEMs face a far smaller market for vehicles which will threaten their existence. The good news is that they have a great opportunity to replace lost profits from vehicles with profits from digital vehicle services. However, to achieve this they must control the app store in their vehicles and remain digitally relevant. This will be no small feat as both Apple and Google will be more than happy to turn them into smartphones on wheels. Digital oblivion will be the price of failure.

• Substantial decline. The main driver of vehicle demand is miles driven. EVs can drive 2.5x more miles than a petrol vehicle before being replaced meaning that without a dramatic increase in miles driven, demand for vehicles could fall by 60% over the next 20 years.
• No fix in sight. With miles being much cheaper to drive, demand should increase materially but without autonomy, this is very unlikely. This is because road congestion will prevent miles driven by humans from increasing. Even with a full roll-out of autonomy driving a 40% increase in miles driven, vehicle demand would still be 30% less in 2047 than it was in 2023.
• OEMs need digital services to survive. By 2040 digital services in the vehicle could be worth $1.7tn while digital advertising could be worth around $61bn. Advertising is not big enough meaning that the OEMs must access digital vehicle services to make ends meet.
• Business model. There is a high likelihood that the business model of the smartphone will be replicated in the digital vehicle. There are two revenue streams. First, the monetisation of the ecosystem and secondly the share of revenues earned from distributing apps and services.
• The digital ecosystems view the vehicle as just another device to be digitised as this gives them both a new revenue opportunity as well as greater stickiness for their ecosystems in terms of running across multiple devices. They will be more than happy to turn OEMs into smartphones on wheels.
• Google is the least threatening of the two major Western ecosystems. This is because it will pay a share of the digital advertising revenue that it earns in the vehicle to the OEM in terms of traffic acquisition cost. The problem is that this will be less than 25% of what OEMs need.
• Apple’s business model threatens to turn OEMs into no more than app developers. Should Apple take over the infotainment unit, then the OEMs will receive nothing.
• The app store is the key to success for OEMs as they must control app and service distribution to vehicle users to earn the revenue share that they need to offset falling vehicle sales. To have a relevant app store, they must remain digitally relevant in terms of the user experience. Using Android without Google services as the OS for the infotainment unit remains their best option for both the user experience and the app store.

RFM research subscribers will receive their copy by email. 

The bubble remains inflated but there are signs that generative AI is not taking over as had been forecast. The economics of AI remain very good even if prices fall by 92% but this means the industry would be 1/12^th of the size previously envisaged. This combined with ongoing problems with AI’s reliability and veracity is likely to cause the bubble to pop and reset expectations to reality.

• AI Bubble: There is very little doubt that AI is currently in a bubble as super-intelligent machines are as far off today as they were 10 years ago. However, generative AI does offer genuine improvements in abilities that have a large number of revenue-generating use cases.
• Usage & search: The first sign is that usage of ChatGPT has stopped growing and it is having no impact on the search market. This is not the trajectory of a revolution and disruption.
• AI economics: $20 per user/month is the current benchmark and at this price, a lot of money will be made. However, RFM thinks that the ever-increasing number of models will mean that the price becomes something closer to $20 per user/year, a decline of 92%. At this price monetisation via advertising becomes a realistic option and many may choose this route.
• The pin: Price erosion of 92% is likely to collapse the business plans of most generative AI start-ups meaning that they will need to raise more money after having missed their targets. This will trigger down rounds and falling valuations causing the bubble to deflate. This will look very similar to what happened with autonomous driving and bike sharing.
• Inference at the edge: is a no-brainer in RFM’s opinion as the economics are far better for the service provider. This is because the service provider does not have to support inference as it is executed on the user’s device.
• Privacy and security: inference at the edge also offers better security and privacy options which are likely to become crucial in the provision of LLM-powered services.
• Customisation: Running models on devices also raises the possibility of substantial customisation where a model on the user’s device is finely tuned to that user’s requirements and preferences.
• Black box & entanglement: are old problems with deep learning and are becoming real issues with generative AI. This is because no one really knows how generative AI does what it does or can promise that it won’t go crazy. This causes real problems when regulators are demanding verifiability and guarantees that models won’t go off the rails.
• Deflation: There is a large revenue opportunity to be had but expectations and valuations need to be adjusted downwards as super-intelligent AI looks to be as far away now as it was 10 years ago.

RFM research subscribers will receive their copy by email. 

There is concern that Microsoft and Google have already lost their edge in AI. The free availability of Meta’s LlaMa models has triggered a wave of innovation in the open-source community. Open-source models are available that are close in performance to those being offered by the big players implying that the lower end of generative AI quickly commoditises. This means that the moats that Google and Microsoft have already managed to create will become key differentiators and RFM thinks that they will protect them while they work out how to adapt to open-source competition.

• Genie escaped. Until recently, building LLMs was only possible for those with lots of cloud compute resources and big bank balances. However, the “leak” of Meta’s foundation model to the market has enabled open source to begin tinkering from which there has been a steady flow of results. This is what underpins fears that the leaders in AI have already lost their edge.
• Self-destructive innovation. The irony is that it is innovations from both Microsoft and Google that have enabled the open-source community to break the dependence on the cloud and begin fine-tuning LLMs at the edge. This has resulted in highly capable models being made available for free to anyone who wants them.
• Innovations: LoRa & Chinchilla. LoRa is a technique invented by Microsoft that enables LLMs to be trained on devices with a fraction of the storage or compute power that was previously required. Chinchilla was invented by Google and is a tweak to training methodology that allows smaller models to perform better than others many times their size.
• Galloping commoditisation. The combination of a freely available foundation model and its weights with these new techniques is how open source is managing to train models on Macs, laptops and PCs with powerful graphics cards. This means that the basic end of this market is going to commoditise extremely quickly meaning that other factors such as developer relationships and user numbers will be extremely important in determining winners and losers.
• Moats. RFM finds that both Google and Microsoft have substantial moats that will protect them from competition from the open-source community. These moats have been created from the fact that both already have well over 1bn users and the fact that ChatGPT is becoming a development platform in its own right. Meta is building its own moat in the open-source community.
• Machiavellian Meta. RFM is convinced that Meta’s “leak” of its LlaMa foundation models and their weights was no accident. “Leaking” LlaMa to open source has meant that LlaMa is now the de facto standard for open-source development and a platform in its own right. These innovations will all be available to Meta to use in its products and services meaning that it has effectively outsourced R&D at almost zero cost.

RFM research subscribers will receive their copy by email. 

The 4^th hype cycle is in full swing but large language models (LLMs) do have a lot to offer. They won’t deliver superhuman intelligence, but they can greatly enhance the productivity of humans as well as their ability to use language as a user interface. These systems are very expensive to develop which is probably a major reason why it is the large companies like Microsoft (OpenAI) and Google that currently lead the race. The AI ecosystem is at a very early stage and remains dominated by Nvidia but for how long remains to be seen.

• 4^th hype cycle. Generative AI has for the 4^th time in 60 years triggered dreams of super-intelligent machines being just around the corner. Generative AI offers a big step forward in terms of ingesting, parsing and retrieving data but it continues to suffer from problems that are very unlikely to be eradicated in the foreseeable future. Hence when superintelligence fails to materialise, there is going to be another round of despondency and depression.
• No breakthrough. The techniques to create generative AI are not new and mostly stem from Google’s 2017 transformer neural network architecture. In 2012, deep learning suddenly started to work thanks to enough data and processing power becoming available. Generative AI looks like a repeat of 2012 rather than a genuine breakthrough that will change the world.
• Use case 1: Enterprise. RFM’s testing reveals that LLMs are extremely good at ingesting huge amounts of data and being able to usefully collate and amalgamate data to produce useful and easy-to-use answers to inquiries. This has substantial use cases within the enterprise and may even render some current corporate data solutions obsolete.
• Use case 2: Automotive: where the current smartphone-like user interface is woefully inadequate. Voice has always been the leading contender but has not been anything like good enough. Generative AI could substantially improve the ability to control vehicles with voice and offers a way back for OEMs to reclaim some of the ground that they have already ceded.
• Generative AI is expensive but right now no one seems to care. RFM finds that for services with millions of users, the cost of inference will be far greater than the cost of training. This creates a substantial advantage in running inference in end devices rather than in the cloud.
• Generative AI players: Despite its recent gaffes, Google is a leader in the generative AI space alongside Microsoft (OpenAI) having created a lot of the innovation that everyone is now using. Meta Platforms, Baidu and Midjouney make up 2^nd place with others like Alibaba, Amazon and SoundHound nipping at their heels. Apple & Samsung are nowhere to be seen.
• The AI ecosystem is at a very early stage and is currently dominated by Nvidia which has 85%+ share of the platform used to train AIs. However, the popularity of ChatGPT could move the control point away from Nvidia and greatly weaken its position.

RFM research subscribers will receive their copy by email. 

Despite all indications to the contrary, China has abruptly changed course by ending Covid Zero and working to get the economy back on track. There are plenty of reasons to doubt a long-term turnaround, but we look set for a big bounce over the next year or two. At the same time, the USA has managed to get The Netherlands and Japan to increase restrictions, but these may be restricted to military applications as opposed to more broadly as the USA desires. China will set its own standards wherever it can and the result will be fragmentation and lower long-term growth for all.

• Damascene moment: President Xi has abruptly changed course by repealing Covid Zero and moving to get the economy back on track. Concern about the economy was almost certainly the main catalyst though civil protest also played a part.
• China’s return: Vice-Premier Liu He has been on a charm offensive at Davos pushing economic recovery and reforms. He was met with scepticism but in the short to medium term, we think that economic stimulus in Q2 2023 will be followed by a sustained economic recovery.
• Middle-income trap. Demographic data, productivity and high debt levels may push China into the middle-income trap which could see GDP growth averaging just 3-4%. This would be very bad news for its ambition to rival the USA as an economic and technological world power.
• No unilateral control: The USA is unable to control the semiconductor equipment market at 10nm to 18nm unlike it can at the leading edge. These nodes use Deep Ultraviolet (DUV) where equipment can be sourced from The Netherlands or Japan. This is why the USA has been negotiating with these countries feverishly over the last few months.
• Netherlands and Japan: on January 28^th 2023, The Netherlands and Japan agreed to increase restrictions on China but we believe that these two countries will limit themselves to targeting technologies that can be used for military purposes as opposed to technology more broadly.
• The new restrictions will 1) affect the non-leading edge nodes (18nm – 10nm), 2) be delayed in implementation, 3) be targeted to military applications as opposed to the USA which targets everything and 4) may also target compound semiconductors.
• Lagging reality. Although there has been watering down from Japan and The Netherlands, there are other actions the USA can take. Many Chinese companies are already abandoning the advanced nodes and moving to 28nm – 45nm where there is still plenty of business.
• Build the wall 2.0. The Balkanisation of the internet looks more certain than ever. The new restrictions will only make China push harder for self-reliance, meaning that it will create its own standards wherever it can. The splitting of the global network into two pieces inevitably means less growth for all of the technology sector in the long term.

RFM research subcribers will receive their copy by email. 

Software is an existential issue for OEMs. Fragmentation and a poor user experience are the most pressing problems that need to be dealt with if OEMs are to be relevant in the digital vehicle. The execution of the digital ecosystems is currently far superior meaning that they are winning the hearts and minds of users. OEMs need to reverse this trend or face becoming commodity box-shifters enabling the digital ecosystems to suck almost all the value out of yet another industry.

• Life or death dilemma. Vehicle demand could halve over the next 20 years leaving OEMs needing new sources of revenue. RFM thinks that digital services are the best option, but this requires OEMs to achieve digital relevance in the vehicle. Software is the key.
• ADAS & Digital cockpit are the two domains that really matter as these are the two segments of the vehicle that interface directly with the user. Hence, the battle for the digital ecosystem will be fought almost entirely in software and these two domains.
• Frankenstein’s monster. The complexity of the vehicle and the OEMs’ desire to retain control raises the likelihood that vehicle software becomes very fragmented. This is terrible news for the user experience and the ability of 3^rd parties to bring their services into vehicles.
• User experience: No one offers a great user experience for the vehicle but crucially, Apple, Google and Tesla are a lot better than the OEMs. This is a serious problem that if not rectified will result in OEMs becoming app developers in their own vehicles and little more.
• Industry consortia are supposed to be one route for OEMs to create Linux-based software that is common to all vehicles and harmonise how data is exchanged. However, RFM finds that for IVI software, the consortia are facilitating fragmentation rather than preventing it.
• Apple & Google. This leaves the way open for Apple and Google to bring a consistent user experience to the vehicle and make it easy for developers to address multiple vehicles. OEMs’ failure to execute on this is why drivers constantly express a preference for Apple CarPlay or Android Auto as the user experience in the vehicle.
• Android (without Google) is by far the best option for the OEMs to compete with the digital ecosystems as Android will at least bring greater software consistency and make it much easier for app developers to get their apps into the vehicles. This will relieve the OEMs of one of their most pressing problems freeing them to focus on the user experience.
• OEMs are not in good shape as many are either not using Android or are struggling to execute on it. This leaves the door open for the digital ecosystems to take over the digital vehicle leaving OEMs with everything to do in a rapidly closing window.