When OpenAI released ChatGPT, it became obvious – and very fast – that we were entering a new era of AI. Every tech company scrambled to release a comparable service or to infuse their products with some form of GenAI. Microsoft, piggybacking on its investment in OpenAI was the fastest to market with impressive text and image generation for the mainstream. Copilot is now embedded across its software, including Microsoft 365, Teams, GitHub, and Dynamics to supercharge the productivity of developers and knowledge workers. However, the race is on – AWS and Google are actively developing their own GenAI capabilities.
AWS Catches Up as Enterprise Gains Importance
Without a consumer-facing AI assistant, AWS was less visible during the early stages of the GenAI boom. They have since rectified this with a USD 4B investment into Anthropic, the makers of Claude. This partnership will benefit both Amazon and Anthropic, bringing the Claude 3 family of models to enterprise customers, hosted on AWS infrastructure.
As GenAI quickly emerges from shadow IT to an enterprise-grade tool, AWS is catching up by capitalising on their position as cloud leader. Many organisations view AWS as a strategic partner, already housing their data, powering critical applications, and providing an environment that developers are accustomed to. The ability to augment models with private data already residing in AWS data repositories will make it an attractive GenAI partner.
AWS has announced the general availability of Amazon Q, their suite of GenAI tools aimed at developers and businesses. Amazon Q Developer expands on what was launched as Code Whisperer last year. It helps developers accelerate the process of building, testing, and troubleshooting code, allowing them to focus on higher-value work. The tool, which can directly integrate with a developer’s chosen IDE, uses NLP to develop new functions, modernise legacy code, write security tests, and explain code.
Amazon Q Business is an AI assistant that can safely ingest an organisation’s internal data and connect with popular applications, such as Amazon S3, Salesforce, Microsoft Exchange, Slack, ServiceNow, and Jira. Access controls can be implemented to ensure data is only shared with authorised users. It leverages AWS’s visualisation tool, QuickSight, to summarise findings. It also integrates directly with applications like Slack, allowing users to query it directly.
Going a step further, Amazon Q Apps (in preview) allows employees to build their own lightweight GenAI apps using natural language. These employee-created apps can then be published to an enterprise’s app library for broader use. This no-code approach to development and deployment is part of a drive to use AI to increase productivity across lines of business.
AWS continues to expand on Bedrock, their managed service providing access to foundational models from companies like Mistral AI, Stability AI, Meta, and Anthropic. The service also allows customers to bring their own model in cases where they have already pre-trained their own LLM. Once a model is selected, organisations can extend its knowledge base using Retrieval-Augmented Generation (RAG) to privately access proprietary data. Models can also be refined over time to improve results and offer personalised experiences for users. Another feature, Agents for Amazon Bedrock, allows multi-step tasks to be performed by invoking APIs or searching knowledge bases.
To address AI safety concerns, Guardrails for Amazon Bedrock is now available to minimise harmful content generation and avoid negative outcomes for users and brands. Contentious topics can be filtered by varying thresholds, and Personally Identifiable Information (PII) can be masked. Enterprise-wide policies can be defined centrally and enforced across multiple Bedrock models.
Google Targeting Creators
Due to the potential impact on their core search business, Google took a measured approach to entering the GenAI field, compared to newer players like OpenAI and Perplexity. The useability of Google’s chatbot, Gemini, has improved significantly since its initial launch under the moniker Bard. Its image generator, however, was pulled earlier this year while it works out how to carefully tread the line between creativity and sensitivity. Based on recent demos though, it plans to target content creators with images (Imagen 3), video generation (Veo), and music (Lyria).
Like Microsoft, Google has seen that GenAI is a natural fit for collaboration and office productivity. Gemini can now assist the sidebar of Workspace apps, like Docs, Sheets, Slides, Drive, Gmail, and Meet. With Google Search already a critical productivity tool for most knowledge workers, it is determined to remain a leader in the GenAI era.
At their recent Cloud Next event, Google announced the Gemini Code Assist, a GenAI-powered development tool that is more robust than its previous offering. Using RAG, it can customise suggestions for developers by accessing an organisation’s private codebase. With a one-million-token large context window, it also has full codebase awareness making it possible to make extensive changes at once.
The Hardware Problem of AI
The demands that GenAI places on compute and memory have created a shortage of AI chips, causing the valuation of GPU giant, NVIDIA, to skyrocket into the trillions of dollars. Though the initial training is most hardware-intensive, its importance will only rise as organisations leverage proprietary data for custom model development. Inferencing is less compute-heavy for early use cases, such as text generation and coding, but will be dwarfed by the needs of image, video, and audio creation.
Realising compute and memory will be a bottleneck, the hyperscalers are looking to solve this constraint by innovating with new chip designs of their own. AWS has custom-built specialised chips – Trainium2 and Inferentia2 – to bring down costs compared to traditional compute instances. Similarly, Microsoft announced the Maia 100, which it developed in conjunction with OpenAI. Google also revealed its 6th-generation tensor processing unit (TPU), Trillium, with significant increase in power efficiency, high bandwidth memory capacity, and peak compute performance.
The Future of the GenAI Landscape
As enterprises gain experience with GenAI, they will look to partner with providers that they can trust. Challenges around data security, governance, lineage, model transparency, and hallucination management will all need to be resolved. Additionally, controlling compute costs will begin to matter as GenAI initiatives start to scale. Enterprises should explore a multi-provider approach and leverage specialised data management vendors to ensure a successful GenAI journey.
GenAI has taken the world by storm, with organisations big and small eager to pilot use cases for automation and productivity boosts. Tech giants like Google, AWS, and Microsoft are offering cloud-based GenAI tools, but the demand is straining current infrastructure capabilities needed for training and deploying large language models (LLMs) like ChatGPT and Bard.
Understanding the Demand for Chips
The microchip manufacturing process is intricate, involving hundreds of steps and spanning up to four months from design to mass production. The significant expense and lengthy manufacturing process for semiconductor plants have led to global demand surpassing supply. This imbalance affects technology companies, automakers, and other chip users, causing production slowdowns.
Supply chain disruptions, raw material shortages (such as rare earth metals), and geopolitical situations have also had a fair role to play in chip shortages. For example, restrictions by the US on China’s largest chip manufacturer, SMIC, made it harder for them to sell to several organisations with American ties. This triggered a ripple effect, prompting tech vendors to start hoarding hardware, and worsening supply challenges.
As AI advances and organisations start exploring GenAI, specialised AI chips are becoming the need of the hour to meet their immense computing demands. AI chips can include graphics processing units (GPUs), application-specific integrated circuits (ASICs), and field-programmable gate arrays (FPGAs). These specialised AI accelerators can be tens or even thousands of times faster and more efficient than CPUs when it comes to AI workloads.
The surge in GenAI adoption across industries has heightened the demand for improved chip packaging, as advanced AI algorithms require more powerful and specialised hardware. Effective packaging solutions must manage heat and power consumption for optimal performance. TSMC, one of the world’s largest chipmakers, announced a shortage in advanced chip packaging capacity at the end of 2023, that is expected to persist through 2024.
The scarcity of essential hardware, limited manufacturing capacity, and AI packaging shortages have impacted tech providers. Microsoft acknowledged the AI chip crunch as a potential risk factor in their 2023 annual report, emphasising the need to expand data centre locations and server capacity to meet customer demands, particularly for AI services. The chip squeeze has highlighted the dependency of tech giants on semiconductor suppliers. To address this, companies like Amazon and Apple are investing heavily in internal chip design and production, to reduce dependence on large players such as Nvidia – the current leader in AI chip sales.
How are Chipmakers Responding?
NVIDIA, one of the largest manufacturers of GPUs, has been forced to pivot its strategy in response to this shortage. The company has shifted focus towards developing chips specifically designed to handle complex AI workloads, such as the A100 and V100 GPUs. These AI accelerators feature specialised hardware like tensor cores optimised for AI computations, high memory bandwidth, and native support for AI software frameworks.
While this move positions NVIDIA at the forefront of the AI hardware race, experts say that it comes at a significant cost. By reallocating resources towards AI-specific GPUs, the company’s ability to meet the demand for consumer-grade GPUs has been severely impacted. This strategic shift has worsened the ongoing GPU shortage, further straining the market dynamics surrounding GPU availability and demand.
Others like Intel, a stalwart in traditional CPUs, are expanding into AI, edge computing, and autonomous systems. A significant competitor to Intel in high-performance computing, AMD acquired Xilinx to offer integrated solutions combining high-performance central processing units (CPUs) and programmable logic devices.
Global Resolve Key to Address Shortages
Governments worldwide are boosting chip capacity to tackle the semiconductor crisis and fortify supply chains. Initiatives like the CHIPS for America Act and the European Chips Act aim to bolster domestic semiconductor production through investments and incentives. Leading manufacturers like TSMC and Samsung are also expanding production capacities, reflecting a global consensus on self-reliance and supply chain diversification. Asian governments are similarly investing in semiconductor manufacturing to address shortages and enhance their global market presence.
Japan is providing generous government subsidies and incentives to attract major foreign chipmakers such as TSMC, Samsung, and Micron to invest and build advanced semiconductor plants in the country. Subsidies have helped to bring greenfield investments in Japan’s chip sector in recent years. TSMC alone is investing over USD 20 billion to build two cutting-edge plants in Kumamoto by 2027. The government has earmarked around USD 13 billion just in this fiscal year to support the semiconductor industry.
Moreover, Japan’s collaboration with the US and the establishment of Rapidus, a memory chip firm, backed by major corporations, further show its ambitions to revitalise its semiconductor industry. Japan is also looking into advancements in semiconductor materials like silicon carbide (SiC) and gallium nitride (GaN) – crucial for powering electric vehicles, renewable energy systems, and 5G technology.
South Korea. While Taiwan holds the lead in semiconductor manufacturing volume, South Korea dominates the memory chip sector, largely due to Samsung. The country is also spending USD 470 billion over the next 23 years to build the world’s largest semiconductor “mega cluster” covering 21,000 hectares in Gyeonggi Province near Seoul. The ambitious project, a partnership with Samsung and SK Hynix, will centralise and boost self-sufficiency in chip materials and components to 50% by 2030. The mega cluster is South Korea’s bold plan to cement its position as a global semiconductor leader and reduce dependence on the US amidst growing geopolitical tensions.
Vietnam. Vietnam is actively positioning itself to become a major player in the global semiconductor supply chain amid the push to diversify away from China. The Southeast Asian nation is offering tax incentives, investing in training tens of thousands of semiconductor engineers, and encouraging major chip firms like Samsung, Nvidia, and Amkor to set up production facilities and design centres. However, Vietnam faces challenges such as a limited pool of skilled labour, outdated energy infrastructure leading to power shortages in key manufacturing hubs, and competition from other regional players like Taiwan and Singapore that are also vying for semiconductor investments.
The Potential of SLMs in Addressing Infrastructure Challenges
Small language models (SLMs) offer reduced computational requirements compared to larger models, potentially alleviating strain on semiconductor supply chains by deploying on smaller, specialised hardware.
Innovative SLMs like Google’s Gemini Nano and Mistral AI’s Mixtral 8x7B enhance efficiency, running on modest hardware, unlike their larger counterparts. Gemini Nano is integrated into Bard and available on Pixel 8 smartphones, while Mixtral 8x7B supports multiple languages and suits tasks like classification and customer support.
The shift towards smaller AI models can be pivotal to the AI landscape, democratising AI and ensuring accessibility and sustainability. While they may not be able to handle complex tasks as well as LLMs yet, the ability of SLMs to balance model size, compute power, and ethical considerations will shape the future of AI development.