Vertex
Vertex: Decentralized exchange resource for on-chain swaps, liquidity, and trading.
ABAB Structured Brief
Vertex is indexed in ABAB Crypto Map under Decentralized Exchanges. This page keeps the official site, category, tags, and related ABAB coverage together as a searchable crypto project profile. Official domain: vertexprotocol.com.
Related News & Analysis
Google Announces Gemini Embedding 2 Now Fully Available in Gemini API and Vertex AI
...ni Embedding 2 is now fully available in the Gemini API and Vertex AI. This is its first native multimodal embedding model, supporting unified representation of various data types such as text and images, and has been op...
From OpenSea to OpenRouter: Alex Atala Analyzes Multi-Model Paradigms, Jevons Paradox, and Dynamic Cost Control for Enterprises
"OpenRouter CEO: Why Chinese Open Models Are Beating the US Why Enterprises Fear OpenAI & Anthropic" (20VC interview with Harry Stebbings, featuring OpenRouter co-founder and CEO Alex Atala), here are the key points summarized: 1. From OpenSea to OpenRouter: High-Concurrency Architecture and Market Evolution • Lessons from OpenSea: Alex was a co-founder of the NFT trading platform OpenSea. Early on, OpenSea experienced massive traffic surges and server downtime risks. He brought the underlying architecture experience of high concurrency, high availability (Uptime), and elastic scaling to OpenRouter, ensuring stability during model surges or service fluctuations. • Rise of Inference Providers: It was initially thought that model hosting would be monopolized by the three major cloud providers (AWS, Azure, GCP), but in reality, specialized inference providers like Fireworks and Together respond faster and perform better in deploying open-weight models (such as GLM, Kimi, DeepSeek). • Nvidia's ecosystem preference: Nvidia tends to diversify customer concentration by allocating GPU quotas to multiple inference providers, fostering a flourishing ecosystem of underlying computing power providers. 2. Multi-Model Future and AI Neurodiversity • Rejecting single-model monopoly: Advocating for "AI Neurodiversity," firmly believing that the future will not be dominated by a single model. Both enterprises and individuals need to use a combination of different models to achieve higher creativity and cost-effectiveness. • Specialization and brand intelligence: Enterprises will not rely solely on a generic model in the future but will fine-tune or train proprietary models (such as using LoRA plugins) for their core business while also utilizing other excellent open-source/closed-source models across the network. • Jevons Paradox validation: Taking GPT-5.6 / Luna as an example, after OpenAI reduced its price by 10 times, usage on the OpenRouter platform surged by 13 times. Lowering model prices does not reduce total expenditure; instead, it exponentially stimulates a larger demand for calls. 3. Why Enterprises Remain Cautious of Closed-Source Giants like OpenAI & Anthropic • Preventing vertical encroachment by giants (e.g., Claude Design vs. Figma): Model vendors have strong incentives to enter vertical application scenarios (e.g., Anthropic launching Claude Design). Enterprises worry that direct ties to closed-source giants will lead to opaque data policies, binding risks, and potential vertical replacement by the giants. • Data risks and VPC needs: Many enterprises find it difficult to fully trust closed-source vendors' data retention and privacy policies, preferring to deploy open-weight models in their own VPC (Virtual Private Cloud) or through open gateways for greater control. 4. The Competition of Open-Source Models Between China and the US: The US is Lagging • Strong momentum of Chinese open-source models: In the open-weight domain, Chinese open-source models (such as DeepSeek, GLM 5.2, Kimi/Moonshot, Qwen, etc.) have made significant breakthroughs in performance, inference efficiency, and writing capabilities. The US is currently lagging in the open-source model field. • Developer usage preferences: In the OpenRouter's ranking of open-source/open-weight model usage, Chinese open-source models have long occupied the top positions. • Distillation and catch-up strategies: Distillation is a conventional scientific method to enhance model efficiency. US Neolabs (new large model laboratories, such as Poolside, Thinking Machines) can quickly catch up through compliant distillation and reinforcement learning (RL), provided they solve the barriers to acquiring computing power. 5. Harness, Agent Architecture, and New Paradigms in Enterprise Management • Difference between Harness and Apps: Harness is built on Unix/command line principles as an Agent control layer, which is more composable, deterministic, and model-friendly than traditional API or UI-based Apps. • Orchestrator and Sub-Agent architecture: The mainstream architecture of the future will be a high-IQ "main orchestration model" coordinating the overall situation, issuing instructions to multiple low-cost, high-deterministic "open-source sub-agents" to execute standardized tasks such as classification and extraction. • Dynamic Employee Cost: Enterprise management will undergo transformation in the AI era. The inference costs incurred by employees using different models are highly dynamic, and in the future, enterprises will need to manage performance and costs based on the match between "employee output" and "AI computing power consumption costs."
Gemini Empire: How Google Rebuilt Its AI Machine
Background and starting point. Gemini was not born as a single isolated model project. It emerged after Google compressed years of work across DeepMind, Google Brain, Google Research, Cloud TPU, Search, Android, and Workspace into one coordinated line of execution. In April 2023, Google merged DeepMind and the Brain team from Google Research into Google DeepMind. Sundar Pichai said the new unit was meant to build more capable general AI systems faster, more safely, and more responsibly, and he explicitly said Jeff Dean would help lead a series of powerful multimodal models. That was the organizational starting point of Gemini. When Gemini 1.0 launched in December 2023, Google described it as the first realization of the vision behind Google DeepMind and one of the biggest science and engineering efforts in the company’s history. Publicly, Google framed the move around capability and safety; in industry context, it also clearly reflected competitive pressure from the OpenAI / Microsoft wave. At the same time, some core facts remain undisclosed: precise parameter counts, full training-token totals, exact modality mix, and true training cost are still publicly limited or unconfirmed. Organization, talent, and governance. Gemini was built through a dual leadership structure centered on Demis Hassabis and Jeff Dean. Hassabis became CEO of Google DeepMind and led the company’s most capable and general AI systems; Jeff Dean became Chief Scientist across Google Research and Google DeepMind, with multimodal model work named as one of his first major strategic assignments. This makes Gemini neither a pure DeepMind-only model nor a pure Google Research-only model. It was the first flagship model family of the merged Google AI organization. Google DeepMind also made clear internally that this would not be a “research island”: it was meant to work closely with Google product areas so that research could be turned into products across Google and Alphabet. In 2024, Google moved Responsible AI teams closer to DeepMind, bringing governance and model-building physically and organizationally nearer together. Over time, the public author lists around Gemini also became more stable, with figures such as Koray Kavukcuoglu, Jeff Dean, Oriol Vinyals, and Noam Shazeer appearing directly in model launches, showing that Gemini was becoming a durable product-and-research line rather than a one-off executive initiative. Technical foundation and engineering base. Gemini sits on several research streams rather than one direct ancestor. Google’s Pathways vision pushed toward a single system that could generalize across many tasks and modalities; PaLM proved Google could train a giant Pathways-based language model across 6,144 TPU v4 chips; DeepMind’s Chinchilla shifted thinking toward compute-optimal training; Flamingo showed that interleaved image / video / text prompting could deliver strong multimodal few-shot behavior; and Gato showed how multiple tasks and modalities could be serialized into one token stream. Gemini then turned this into a native multimodal product family. Google repeatedly said Gemini was trained as a multimodal model from the start, not built by stitching together separate modality systems after the fact. The Gemini technical report shows interleaved text, image, audio, and video inputs, and even interleaved image-and-text outputs. On the infrastructure side, Gemini 1.0 used TPU v4 and v5e, with Ultra trained across a large TPU v4 fleet spanning multiple data centers. Gemini 1.5 used multiple 4,096-chip TPU v4 pods across multiple data centers, and its pretraining data included web documents, code, images, audio, and video, followed by instruction tuning and human-preference tuning. Gemini 2.0 then moved even further into Google’s custom hardware stack, with Google stating that 100% of Gemini 2.0 training and inference ran on Trillium TPUs. Exact data ratios, filtering rules, and licensing proportions remain publicly limited or unconfirmed. How the model family was actually built across generations. Gemini 1.0 launched in December 2023 as Ultra, Pro, and Nano. Google positioned it as its largest and most general model family, claimed state-of-the-art results on most of the benchmarks it reported, and immediately connected it to Bard, Pixel, AI Studio, and Vertex AI. Gemini 1.5 was the real turning point from “strong multimodal model” to “efficient long-context model.” Google said the 1.5 generation reflected research and engineering changes across nearly every part of foundation-model development and infrastructure, especially a new Mixture-of-Experts architecture. The key message was that 1.5 Pro could reach roughly Ultra-level quality with less compute, while pushing context from 128k toward 1 million tokens in product and up to around 10 million tokens in research settings. The later 1.5 technical report described near-perfect retrieval in long-context tasks and strong long-document, long-code, long-video, and long-audio performance. Gemini 2.0 changed the goal again: Google framed it as a model family for the “agentic era,” with native image and audio output, native tool use, and direct use in Project Astra, Project Mariner, Jules, and Deep Research. Gemini 2.5 then became Google’s explicit “thinking model,” and Google said these reasoning capabilities would increasingly be built directly into all its models. By Gemini 3.1 and 3.5, the line had moved further toward long-horizon, agentic workflows and multi-step execution. So the real build story is cumulative: native multimodality, then efficient long context, then tool use, then explicit reasoning, then increasingly agentic execution. Commercialization, distribution, and capital logic. Gemini’s business model is multi-layered. On the consumer side, Google renamed Bard to Gemini in February 2024, launched the Gemini app, and introduced Gemini Advanced through the Google One AI Premium subscription at $19.99 per month. That turned an experimental chatbot into a branded, paid consumer AI line. Over time, the subscription ladder expanded into current Google AI Pro / Ultra-style offerings. On the developer and enterprise side, Gemini became a token-metered platform product through Google AI Studio, the Gemini API, Vertex AI, and Gemini Enterprise. On the distribution side, its power comes less from the standalone app than from Google’s ability to insert Gemini into high-traffic surfaces: Search said AI Overviews were powered by a custom Gemini model; Workspace said Gemini in side panels would use 1.5 Pro; Samsung’s Galaxy S24 became the first major external mobile channel to deploy Gemini Pro and related Gemini capabilities at global consumer scale. On the capital side, Gemini is backed by Alphabet’s infrastructure spending. In Alphabet’s 2025 Q1 earnings call, the company said quarterly CapEx was $17.2 billion, mainly for technical infrastructure, with servers first and data centers second, specifically to support Google Services, Google Cloud, and Google DeepMind; it also maintained an approximately $75 billion full-year CapEx expectation for 2025. That is why Gemini is not just a model family but a Google-scale system. It also serves defensive and offensive business goals: Alphabet said AI Overviews already help drive Search usage, and monetization has remained roughly in line with traditional Search formats. Controversies, limitations, and present position. Gemini’s first major controversy was about demonstration credibility. Shortly after launch, reporting showed that one high-profile Gemini demo video had been edited and did not reflect a fully real-time spoken interaction; later, after scrutiny from the U.S. advertising self-regulator NAD, Google stopped promoting that video. The second major controversy involved image generation of people. In February 2024, Google admitted that the Gemini app’s people-image generation feature, built on top of Imagen 2, had produced inaccurate and sometimes offensive results, especially in historical and cultural contexts, and paused the feature. Google’s own explanation was that diversity-related tuning had been applied too broadly in cases where it should not have been, while the system had also become overly cautious and refused some benign prompts. On safety more broadly, Google has consistently said Gemini undergoes extensive safety evaluation, and DeepMind’s dangerous-capability evaluation program reported no evidence of strong dangerous capabilities in the Gemini models they tested, while still flagging early warning signs. Today, Gemini is no longer just “Google’s answer to ChatGPT”; it is a central AI substrate across the company. At I/O 2026, Sundar Pichai said Google was processing more than 3.2 quadrillion tokens per month across its surfaces, that 8.5 million developers were building monthly with Google’s models, that AI Overviews had passed 2.5 billion monthly active users, and that the Gemini app had surpassed 900 million monthly active users. The most accurate high-confidence conclusion is this: Gemini was built not by one paper or one benchmark win, but by five things happening at once — organizational merger, a native multimodal technical direction, custom TPU infrastructure, product-wide distribution, and steady iteration toward agentic execution. What remains uncertain are exact model sizes, complete data composition, full post-training recipes, and true per-generation training costs; public information on those remains limited or unconfirmed.
Temasek Uncovered: The Rise of Singapore’s State-Owned Capital into a Global Investment Powerhouse
First, the name needs to be clarified. In Chinese usage, “Temasek” is often loosely described as a “fund,” but the more accurate legal and institutional subject here is Temasek Holdings, not a conventional fund or a charitable foundation. Temasek itself states that it was incorporated in 1974 under the Singapore Companies Act, that it owns its assets, and that it is not a fund manager. Singapore’s Ministry of Finance further states that the Government is Temasek’s sole equity shareholder, while Temasek owns the assets on its own balance sheet. In other words, although international media often discuss Temasek as part of Singapore’s sovereign investment system, the more precise description is: a commercially run global investment company wholly owned by Singapore’s Ministry of Finance and subject to constitutional safeguards. If by “Temasek fund” you actually meant Temasek Foundation, that is a different institutional line. Official materials show that since 2003, Temasek has set aside part of its net positive returns above its risk-adjusted cost of capital for community gifts; in 2007, it established Temasek Trust to manage and disburse those endowed funds, and Temasek Foundation to design and deliver programmes. Temasek, Temasek Trust, and Temasek Foundation are connected, but they are not the same legal entity, do not share the same governance structure, and Temasek does not direct the day-to-day operations of the latter two. This report therefore focuses on Temasek Holdings, with the philanthropic ecosystem treated as a related but separate branch. In terms of “founding background,” Temasek is almost a direct institutional product of Singapore’s early industrialisation model. It was created in 1974 to take over a basket of investments and corporate stakes previously held by the Minister for Finance, with an initial portfolio valued at S$354 million. Those assets were not glamorous: Temasek’s own FAQ lists a bird park, a hotel, a shoemaker, a detergent producer, naval yards converted into a ship-repair business, a start-up airline, and an iron and steel mill. The core purpose was to shift the Government away from directly running companies and toward policy-making and regulation, while placing enterprise assets into a commercially run vehicle. Temasek’s history page also notes that of the original 35 companies, 10 still remain directly or indirectly in the portfolio. Its “growth environment” was not a family but the institutional soil of Singaporean state capitalism. Temasek is neither a government department nor a statutory board; it is a company under the Companies Act. At the same time, it is a Fifth Schedule entity under the Singapore Constitution, with a duty to protect its “past reserves.” That gives it a dual mandate: it must operate commercially, yet it cannot freely risk or consume historically accumulated national wealth. Its relationship with the President is part of Singapore’s constitutional “second key” framework: any draw on past reserves requires presidential approval. This design made Temasek, from the start, something very different from a normal private equity firm or a standard state holding company. It became, in effect, an institutional interface for the market-based management of national capital. As of now, Temasek’s latest full public annual disclosure covers the financial year ended 31 March 2025. In July 2025, Temasek reported a net portfolio value of S$434 billion, up S$45 billion from the previous year; if its unlisted portfolio were marked to market, total value would rise to S$469 billion. Its 20-year total shareholder return in Singapore-dollar terms was 7%, its 10-year return 5%, and its since-inception return 14%. Temasek also stated that it had roughly quadrupled its portfolio over the last two decades. In pure institutional terms, that places it among the most successful state investment institutions in the world. Governance, capital structure, and power relations Temasek’s core power structure is not especially complicated, but it is layered. Its sole shareholder is the Singapore Minister for Finance as a corporate body; the shareholder has the right to appoint, reappoint, or remove directors, but those actions require the President’s concurrence, and the appointment or removal of the CEO is also subject to presidential concurrence. The Board is accountable for long-term performance and delegates day-to-day management to the executive team. Temasek explicitly states that there are no nominees of the Singapore Government or any other government on its Board. As of 31 March 2025, the Board had 14 members, of whom 86% were non-executive, independent, and primarily private-sector business leaders. Leadership has now entered a new generational phase. Temasek’s board page shows that Teo Chee Hean became Chairman on 9 October 2025, while Tan Chong Meng became Deputy Chairman on the same day. On the management side, Dilhan Pillay Sandrasegara has served as Executive Director and CEO since 1 October 2021, and after Temasek’s recent organisational restructuring he remains the central operating leader across Temasek and its key wholly owned entities. From a capital-structure perspective, the biggest difference between Temasek and an ordinary fund is that it is not built on a classic LP–GP model. It is an institution with its own balance sheet and its own assets. Temasek says its funding comes mainly from divestment proceeds, dividends from portfolio companies, and distributions from funds it has invested in. Those recurring inflows can be supplemented by bond issuance, euro-commercial paper, bank borrowing, and—where the shareholder chooses—capital injections into Temasek shares. Over the past five years, Temasek averaged about S$32 billion to S$36 billion in annual divestments and about S$9.7 billion to S$10 billion in annual dividends from portfolio companies. The implication is important: Temasek does not rely on perpetual fundraising narratives; it relies on an existing asset base that generates its own cash flows. Its borrowing profile also reflects the advantages of a state-backed investment company with top-tier credit quality. As of 31 March 2025, Temasek’s total debt was only 5% of net portfolio value and 17% of liquid assets. It had 26 Temasek Bonds outstanding, totalling around S$20.2 billion, with a weighted average maturity of more than 18 years. Its corporate ratings were Aaa from Moody’s and AAA from S&P. In practical terms, this means Temasek has deep access to low-cost, long-duration funding without running an aggressive leverage model. Temasek’s relationship with the Singapore budget is also frequently misunderstood. Both Temasek and the Ministry of Finance stress that Singapore’s Net Investment Returns framework allows the Government to spend part of the expected long-term real returns of the investment entities, but that framework does not determine the amount of dividend Temasek must pay in any given year and does not alter Temasek’s independent investment mandate. Temasek continues to declare dividends under its Board-approved dividend policy, balancing current distributions against retained capital for future reinvestment. Put simply: Temasek is very important to the state, but it is not a cash account that the Ministry of Finance can simply draw down at will. Beyond capital, Temasek also has a powerful global network function. It has 13 offices across 9 countries and maintains multiple advisory panels and consultation networks. Officially disclosed advisers include figures such as former World Bank President Robert Zoellick, former Australian Prime Minister Malcolm Turnbull, former Chinese Finance Minister Lou Jiwei, former UBS Chairman Axel Weber, and senior leaders from LVMH, AXA, Mitsui, FPT, and others. These people do not necessarily execute deals for Temasek, but they form an important layer of influence capital and information-network capital around the institution. Asset map and business model As of 31 March 2025, Temasek’s portfolio structure is highly legible. By headquarters exposure, 52% of the portfolio was in Singapore-headquartered companies, 19% in the Americas, 11% in China, 11% in Europe/Middle East/Africa, and 5% in India. But on a look-through underlying-country basis, Singapore was only 27%, China 18%, India 8%, the rest of Asia Pacific 11%, the Americas 24%, and Europe/Middle East/Africa 12%. Temasek also stresses that while it is “anchored in Asia,” it has 66% underlying exposure to developed economies. That is a clear sign that Temasek is no longer simply an owner of Singapore state enterprises; it is now an Asia-rooted, globally allocated investment institution. Sectorally, in 2025 its portfolio consisted of 22% transportation and industrials, 22% financial services, 20% telecommunications/media/technology, 13% consumer and real estate, 7% life sciences and agri-food, 9% multi-sector funds, and about 7% other assets including credit. By liquidity profile, 49% of the portfolio was in unlisted assets and 51% in liquid and listed assets. Temasek estimates that if the unlisted portfolio were marked to market rather than carried at book value less impairment, it would add S$35 billion in value. It also states that the unlisted portfolio generated 7% annual returns over the last decade and more than 10% over the last two decades, outperforming the listed portfolio. This is a crucial point: a growing share of Temasek’s value creation now comes from large-scale private assets, platforms, and long-hold unlisted positions. Its most important hard assets remain its Singapore core holdings. According to Temasek’s official major investments list as of 31 March 2025, it held 28% of DBS, 53% of Singapore Airlines, 51% of Singtel, 51% of ST Engineering, 100% of PSA International, 100% of Singapore Power, 100% of Mapletree, 100% of CapitaLand Group, 100% of SMRT, 100% of Mandai Park Holdings, 52% of Olam Group, 50% of Sembcorp Industries, 40% of SATS, 36% of Seatrium, and 20% of Keppel. These are not merely visible brands; they are direct economic control positions across Singapore’s banking, aviation, port infrastructure, telecoms, power, real estate, transit, engineering, and food systems. Its international book shows how far it has moved from being only a Singapore holding company. Officially disclosed major investments include 17% of Standard Chartered, 5% of Adyen, 3% of AIA, 3% of BlackRock, 19% of VFS Global, 14% of Neoen, 88% of Element Materials Technology, and 33% of Manipal Health, as well as positions in Amazon, Alibaba, Tencent, Visa, Mastercard, NVIDIA, HDFC Bank, ICICI Bank, Axis Bank, Ping An, Meituan, and Eternal, among others. What makes Temasek distinctive is not just that it invests globally, but that it manages strategic control assets and global growth investments on the same balance sheet. Its business model is therefore not a simple “buy and exit” model. In 2025 Temasek divided the overall portfolio into three segments: 41% Singapore-based Temasek Portfolio Companies, which provide stable long-term returns and dividends; 36% Global Direct Investments, aligned with four structural trends—Digitisation, Sustainable Living, Future of Consumption, and Longer Lifespans; and 23% Partnerships, Funds, and Asset Management Companies, which provide access to external partnerships, alternative assets, private credit, diversified fund exposure, and capital-solutions businesses. From 1 April 2026, those three segments are mapped onto Temasek Singapore, Temasek Global Investments, and Temasek Partnership Solutions, while Temasek International retains group and corporate functions. This restructuring says a great deal: Temasek is redesigning itself from one giant holding vehicle into a three-engine capital system. The most interesting platforms under Temasek are not just stock holdings but pieces of capital infrastructure. Temasek says the relevant asset-management platforms together had more than S$90 billion in assets under management as of 31 March 2025, including ABC Impact, Aranda Principal Strategies, Decarbonization Partners, Heliconia, Pavilion Capital, Seviora Holdings, Vertex Holdings, and 65 Equity Partners. Seviora, established in 2020, was launched as a multi-asset asset-management group with approximately S$75 billion of initial AUM. 65 Equity Partners, launched in 2021, started with S$4.5 billion to provide equity and structured capital to established businesses and family-owned firms. GenZero, launched in 2022 with an initial S$5 billion commitment, focuses on global decarbonisation investing. In other words, Temasek does not only buy assets; it builds platforms that institutionalise, productise, and extend the life of capital. On the “influence asset” side, Temasek has three especially strong soft-power pillars. The first is disclosure and institutional reputation: since 2004 it has published the Temasek Review, and its own FAQ states that its reporting standards exceed the baseline of the Santiago Principles. The second is sustainability and climate positioning: as of 31 March 2025, investments aligned to the Sustainable Living trend totalled S$46 billion, and Temasek targets a 50% reduction in attributable portfolio emissions by 2030 from 2010 levels, with a net-zero ambition by 2050. The third is its community ecosystem: Temasek says its gifts to Temasek Trust have supported programmes that have impacted around 4.4 million lives. Together, these form a kind of legitimacy moat around Temasek as a state-owned capital institution. Turning points, controversies, and present-day position If Temasek’s fifty-year history is compressed into a short timeline, the most important moments are these. In 1974, it took over government-held investments and institutionalised the shift away from direct state ownership of enterprises. In the early 2000s, it stepped out of Singapore and built positions in China and India, betting on rising Asia. In the 2010s, it expanded further into the United States and Europe and became a genuinely global investor. In 2020, it created Seviora and platformised more of its asset-management capability. In 2021, Ho Ching retired and Dilhan Pillay took over, completing a major generational management transition. In 2022, it launched GenZero and formalised climate investing as a platform. In 2025–2026, it reorganised around the TSG / TGI / TPS architecture. Each of these moves reflected a shift in Temasek’s reading of the world: from national industrial capital, to Asian growth capital, to global long-horizon capital, and now toward a platform-based, multi-engine, theme-driven investor. Its greatest achievement is not one famous deal but the transformation of a set of Singapore domestic enterprises from nation-building tools into a portfolio base capable of generating durable dividends, global capabilities, and resilience across cycles. Temasek states plainly that its Singapore portfolio companies provide stable long-term returns and dividend income that funds much of its wider investment activity, and that those companies together generate more than S$150 billion in revenue and employ over 160,000 people in Singapore. In other words, Temasek does not invest “outside” the state first and foremost; it first builds a strategic national core and then seeks external growth around it. That is a major reason it is so often treated as a model of state capital deployment. The second reason observers remember Temasek is that it combines strategic holdings, financial returns, global thematic investing, platform incubation, and philanthropic allocation within a relatively coherent institutional framework. Many sovereign investors focus mainly on liquid markets; many state holding companies focus almost entirely on domestic strategic assets; many foundations are fully ring-fenced from commercial capital. Temasek tries to do all of these things at once and to explain them through one charter, one long-term-return philosophy, and one integrating narrative of “Do Well, Do Right, and Do Good.” Whether that model can always succeed is another matter, but the institutional complexity itself is real and distinctive. In recent years, the single biggest reputational controversy has been FTX. Temasek disclosed that between October 2021 and January 2022 it invested US$210 million in FTX International and US$65 million in FTX US, for a total cost of US$275 million, equal to 0.09% of its net portfolio value as of March 2022. Temasek also disclosed that it had conducted about eight months of due diligence, reviewed audited financial statements, and used external legal and cybersecurity review focused on regulation, AML/KYC, sanctions, and other issues. Even so, after the collapse of FTX it wrote the entire investment down to zero. A later chairman’s statement said an independent internal review found no misconduct by the investment team, but that both the team and senior management took collective accountability and had their compensation reduced. The episode did not threaten Temasek’s balance sheet, but it clearly damaged its public image as a highly prudent judge of risk and management quality. A second type of controversy has involved major cyclical misjudgments. Financial media and commentary have long treated Temasek’s Merrill Lynch / Bank of America investment during 2007–2009 as a cautionary tale. The Financial Times wrote in 2025 that Temasek suffered an estimated US$4.6 billion loss on that episode and linked it, together with FTX, to recurring reputational damage. Earlier, in 2006, Temasek’s acquisition of shares in Shin Corp in Thailand was itself conducted through a structured, formally legal transaction, but it later became widely associated with Thai political backlash and social unrest; Reuters’ 2025 review of a Thaksin tax case still linked the transaction to the conflict of that period. In other words, Temasek’s greatest historical risk has not always been the absolute amount of money lost, but the fact that in politically sensitive sectors and politically sensitive countries, capital transactions can be reinterpreted as political events. A third area of dispute concerns how Temasek’s performance should be measured. One critic’s view is that its 5% ten-year return is respectable but not exceptional relative to broad global equity benchmarks; Reuters in 2025 explicitly compared it with roughly 9% over ten years for the MSCI ACWI. The Singapore official view is different: Temasek is not a fund manager, public market indices are not its “direct benchmark,” and its portfolio includes strategic control assets, substantial unlisted holdings, and a more complex role than that of an ordinary investment fund. This debate is unlikely to disappear soon, because it is fundamentally asking a deeper question: should Temasek be judged as a private investment fund, or as a state capital platform? Public materials do not provide a single universally accepted answer. Seen from the vantage point of mid-2026, Temasek’s present-day position is very clear. It remains one of Singapore’s most important state investment institutions and one of the world’s most consequential pools of long-term capital. By the latest official figure, it manages a portfolio of S$434 billion, has 13 global offices, carries Aaa/AAA credit ratings, spans both national strategic assets and global growth assets, and is reorganising itself around a more explicit three-part structure while prioritising AI, infrastructure, climate transition, and alternative assets. Its 2025 decision to join the AI Infrastructure Partnership launched by Microsoft, BlackRock, and MGX further shows that Temasek intends to sit close to the front of the next global capital-expenditure cycle. If the institution had to be summarised in one sentence today, it would be this: Temasek is not simply a Singaporean state holding company, not merely a sovereign wealth fund, and not just another PE/VC platform; it is a state-owned but globally operated, long-duration, hybrid investment institution. Its greatest strength is its ability to combine core national assets, global capital markets, thematic investing, platform-based asset management, and philanthropic legitimacy in one structure. Its greatest challenge is whether, in a world of greater geopolitical friction, technological volatility, and public scrutiny, it can continue to prove that it can make money, preserve discipline, and avoid letting capital misjudgments turn into institutional reputation loss.
From OpenAI to Anthropic: How Dario Amodei Challenged the AI World Order
The core story is that Dario Amodei is not merely “another AI founder.” He was a central figure in the GPT-2 / GPT-3 / RLHF generation of research, and later turned “safe, steerable, interpretable AI” into Anthropic’s organizational philosophy, governance identity, and commercial differentiation. If reduced to one line, his trajectory is this: a San Francisco-born, public-school-educated scientist shaped by mathematics, moral seriousness, and his father’s death moved from theoretical physics into biophysics and neuroscience, then into Baidu, Google Brain, and OpenAI, and finally built Anthropic as a company that combines frontier-model development, governance design, enterprise distribution, and a safety-centered brand. Dario Amodei was born in San Francisco in 1983 and grew up in the Mission District with his younger sister Daniela. Their father, Riccardo Amodei, was an Italian leather craftsman; their mother, Elena Engel, managed library renovation and construction projects. Public material does not establish a precise wealth class, but the family appears—based on occupations, schooling, and biographical descriptions—to have been rich in educational and civic-cultural capital rather than venture or startup capital. Interviews describe Dario as a child obsessed with numbers and mathematics, and Amodei himself has said his parents gave him a strong sense of right and wrong. He attended Lowell High School, made the 2000 U.S. Physics Olympiad team, studied physics at Caltech before transferring to Stanford, and then completed a Princeton Ph.D. in physics/biophysics focused on neural circuits, later receiving the Hertz Thesis Prize. After Princeton he became a postdoctoral scholar at Stanford Medicine, working on biomedical and proteomic problems. His father’s death in 2006 from a rare illness was a major turning point: Amodei has repeatedly said this experience made him intensely aware of how a few years of scientific acceleration can mean life or death. His first major industry role came in 2014 at Baidu, after Andrew Ng recruited him into work related to speech systems. Public interviews suggest this was where he first developed a strong intuition for scaling: more data, larger models, and longer training meaningfully improved model performance. He then moved to Google Brain as a senior research scientist and joined OpenAI in 2016. Official and near-official sources agree that at OpenAI he became Vice President of Research, helped lead GPT-2 and GPT-3, co-led research direction with Ilya Sutskever, and is credited on his personal site as a co-inventor of RLHF. That matters because it places him directly inside the main capability pipeline of large language models, not merely on the governance or communications side. Anthropic was founded in 2021 as a Delaware Public Benefit Corporation. The precise full founder list is reported inconsistently across public sources, so the most reliable statement is that Dario Amodei and Daniela Amodei are the central co-founders, serving as CEO and President respectively, and that the company emerged from a group of former OpenAI insiders. Before Anthropic became known for Claude, it became known for a research posture: papers such as Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback and Constitutional AI: Harmlessness from AI Feedback made “helpful, honest, harmless” and “Constitutional AI” core parts of the company’s identity. Anthropic reportedly had an early Claude system trained by summer 2022 but delayed broader commercialization for additional internal safety testing; Claude was then formally introduced in March 2023. This decision became a defining part of Anthropic’s reputation as the company willing to trade speed for safety signaling, even though it cost consumer mindshare against ChatGPT. By 2025–2026, Claude had evolved from a single assistant into a product stack. Official Anthropic pages list Claude, Claude Code, Claude Code Enterprise, Claude Cowork, Claude Security, and integrations for Chrome, Slack, and Microsoft 365, along with Opus, Sonnet, Haiku, and Mythos Preview model lines. Claude 4 launched in May 2025 with Opus 4 and Sonnet 4, while Claude Code entered general availability; by late May 2026, official docs and release notes identify Claude Opus 4.8 as the most capable generally available Claude model, with a default 1 million token context window. Anthropic’s assets now fall into two buckets: commercial assets such as the model APIs, subscriptions, enterprise plans, and cloud distribution; and influence assets such as Claude’s Constitution, the Responsible Scaling Policy, the Long-Term Benefit Trust, the Anthropic Institute, the Transparency Hub, the Economic Index, and Project Glasswing. The latter do not merely decorate the company—they function as governance and legitimacy infrastructure. Anthropic’s governance model is one of its strongest differentiators. The Long-Term Benefit Trust is designed as an independent body that will eventually gain the power to select a majority of the board, with the explicit goal of aligning the company with “the long-term benefit of humanity” rather than only shareholder returns. Current board and trust structures are publicly listed by Anthropic. In practice, whether this structure can fully counteract capital pressure remains an open question, but it unquestionably turns governance into part of the company’s public product. This is reinforced by the Responsible Scaling Policy, by the Anthropic Institute launched in March 2026 under Jack Clark, and by initiatives such as Project Glasswing, which tied Anthropic to major firms and institutions in critical software and cyber defense. Anthropic’s capital structure shows that it is not an outsider startup. It raised $580 million in Series B in 2022, officially led by Sam Bankman-Fried, followed by a $450 million Series C in 2023 led by Spark with participation from Google and others. Amazon committed up to $4 billion beginning in 2023 and completed that investment in 2024; by April 2026, Anthropic announced an expanded Amazon relationship involving over $100 billion in AWS technology commitments over ten years, up to 5 gigawatts of compute, and a new $5 billion investment with up to $20 billion more possible. Google Cloud had already become an early preferred cloud partner in 2023; Reuters later reported that Alphabet would invest up to $40 billion in Anthropic and that Anthropic had committed to spend $200 billion on Google Cloud over five years. Microsoft and NVIDIA also announced major strategic investments in 2025, while Claude was made available across AWS, Vertex AI, and Microsoft Foundry. This means Anthropic has built a rare position: deeply tied to all major cloud ecosystems without being wholly captive to one. The company’s business model is unusually explicit. Anthropic stated in 2026 that it makes money through enterprise contracts and paid subscriptions, not advertising, and reinvests that revenue in Claude. Reuters has also reported that the company sells access both directly and through third-party cloud services. By 2025–2026, that translated into a multilayered revenue stack: subscriptions, seat-based team plans, enterprise access fees, API usage, cloud marketplace sales, and vertical solutions. Financially, the growth has been extraordinary: Reuters reported annualized revenue of about $875 million in early 2025; Anthropic later said its run-rate was about $9 billion by the end of 2025, above $30 billion by April 2026, and above $47 billion in May 2026. Those are run-rate figures rather than a single audited annual revenue number, so they should be interpreted with care, but they still show that Anthropic has become one of the fastest-growing AI businesses in the world. Dario Amodei’s biggest strengths are not confined to one paper or one product. He is remembered because he successfully combined three roles that are usually separate: frontier-model builder, safety-governance spokesperson, and founder-CEO capable of translating that identity into enormous capital partnerships and enterprise adoption. At the same time, he faces persistent criticism. Some argue Anthropic’s safety rhetoric coexists with aggressive scaling and fundraising; that criticism sharpened when RSP 3.0 no longer preserved Anthropic’s earlier hardest unilateral “pause if necessary” framing. Others point to the company’s copyright disputes: Reuters reported ongoing music-publisher litigation, additional publisher suits in 2026, and a $1.5 billion proposed settlement in a books case. There is also policy criticism: Anthropic has been accused by opponents of fear-based regulatory capture, even as the company presents itself as unusually transparent and safety-conscious. The Pentagon dispute in 2026 crystallized Anthropic’s real-world position. In Dario Amodei’s official statement, Anthropic said it had already deployed models in classified U.S. government networks, in the national labs, and broadly across military and intelligence work. Yet it refused to remove guardrails against two uses: mass domestic surveillance and fully autonomous weapons under current reliability conditions. Reuters reported that this refusal escalated into a major confrontation with the U.S. defense establishment. This is perhaps the clearest picture of both Amodei and Anthropic: not anti-state, not anti-power, not anti-acceleration—but trying to set boundaries inside an acceleration race they are absolutely still participating in. As of late May 2026, Dario Amodei’s real position is no longer that of “former OpenAI executive.” He is now one of the tiny number of people who can shape frontier-model design, enterprise buying decisions, cloud-provider strategy, AI-safety discourse, and national-security boundaries at the same time. Anthropic continues expanding internationally, lists multiple European offices, and has built institutions such as the Anthropic Institute, the Economic Index, and Project Glasswing to extend its role beyond products into policy and social interpretation. The most accurate conclusion is not that Amodei is simply “the conscience of AI,” nor that Anthropic is merely “OpenAI with better safety marketing,” but that he has helped build one of the most consequential attempts to make frontier AI simultaneously powerful, commercially dominant, governable, and socially legible. Open questions and limitations. Public sources remain incomplete on several points: the exact full founder list is inconsistent across sources; the family’s precise economic class is not formally documented; the exact disease that caused Riccardo Amodei’s death is not reliably confirmed in the most authoritative public material; internal details of Dario’s split from OpenAI are only partially public; and private-company cap-table details shift rapidly and should not be treated as fixed facts. Those gaps matter, and where they exist, the careful answer is not certainty but restraint.
Anthropic Releases Claude Opus 4.7
...orm, API, and cloud services such as Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, with pricing remaining consistent with the previous generation. Source: Public Information