The money doubled. The number of companies getting it did not.

The doubling of private investment — the biggest year-over-year jump since 2014 — happened while the pool of recipients kept shrinking. The number of newly funded AI companies fell from 1,051 in 2019 to 762 in 2020 and 746 in 2021, the third year of a decline that began in 2018. The average private investment deal in 2021 was 81.1% larger than in 2020. Money did not spread out; it concentrated.

Bigger rounds, not more of them

• Rounds worth $500 million or more went from 4 in 2020 to 15 in 2021 — rounds over $1 billion rose from 3 to 5, and the $500 million to $1 billion band from 1 to 10.
• The $100 million to $500 million band more than doubled, from 93 rounds to 235, and the $50 million to $100 million band more than doubled as well, from 85 to 194.
• Rounds under $50 million barely moved: 2,102 in 2020, 2,120 in 2021. All the growth came from the top of the range.
• Across every band, disclosed and undisclosed, total funding rounds rose from 2,638 to 2,959.
• The largest private investments of 2021 went to two data management companies and two robotics and autonomous driving companies — Databricks in the United States, Beijing Horizon Robotics in China, Oxbotica in the United Kingdom and Celonis in Germany.

Data infrastructure overtakes healthcare

• “Data management, processing, and cloud” drew the most private AI investment in 2021 at around $12.2 billion — 2.6 times the roughly $4.69 billion it took in 2020.
• Medical and healthcare came second at $11.29 billion, followed by fintech ($10.26 billion), autonomous vehicles ($8.09 billion) and semiconductors ($6.0 billion).
• Over the five years from 2017 to 2021 the ranking is different: medical and healthcare leads with $28.92 billion, then data management, processing and cloud ($26.91 billion), fintech ($24.92 billion) and retail ($21.95 billion). The 2021 result is a change of direction, not a long-standing lead.
• The categories with a steady multiyear climb were autonomous vehicles, cybersecurity and data protection, fitness and wellness, medical and healthcare, and semiconductors.

One country, three times over

US private AI investment reached $52.88 billion in 2021 — more than three times China’s $17.21 billion and over eleven times the United Kingdom’s $4.65 billion, with Israel ($2.41 billion) and Germany ($1.98 billion) next. Measured against regions rather than countries, the US led China by 3.1 times and the European Union ($6.42 billion) by 8.2 times, and all three grew from 2020. Aggregated across 2013–2021 the ordering is unchanged: the United States $149.0 billion, China $61.9 billion, then the United Kingdom and India at $10.8 billion each and Israel at $6.1 billion. The largest acquisitions of the year sat mostly in health and security — Microsoft bought Nuance Communications for $19.8 billion, Siemens bought Varian Medical Systems for $17.2 billion, Thoma Bravo bought Proofpoint for $12.4 billion, and NortonLifeLock bought the Czech company Avast for $8.0 billion.

• Startup counts follow the same map. In 2021 the United States produced 299 newly funded AI companies, against China’s 119, the United Kingdom’s 49 and Israel’s 28 — roughly two and a half times China on the year.
• Aggregated from 2013 to 2021 the gap is wider still: 3,234 newly funded US companies against 940 in China and 427 in the United Kingdom.
• The decline in newly funded companies is not a global-South story. It shows up in the United States after 2018 and in China after 2019 — the two largest markets are the ones thinning out.

The more interesting detail is the second derivative. Most countries and regions in the index saw their AI hiring growth slow between 2020 and 2021 — the pace at which AI hiring outruns general hiring declined almost everywhere. Germany and Sweden were the exceptions.

Where the postings are

• As a share of all job postings in 2021, AI roles were most common in Singapore (2.33%), then the United States (0.90%), Canada (0.78%), the United Kingdom (0.74%), Australia (0.58%) and New Zealand (0.25%).
• That share rose in the United States, Canada, Australia and New Zealand from 2020 to 2021, and fell in Singapore and the United Kingdom.
• Inside the United States, machine learning is the largest skill cluster at 0.57% of all postings, followed by artificial intelligence (0.33%), neural networks (0.15%), natural language processing (0.13%), robotics (0.11%), visual image recognition (0.10%) and autonomous driving (0.06%).
• Machine learning postings stand at nearly three times their 2018 level and artificial intelligence postings at around 1.5 times, despite small dips in both from 2019 to 2020.
• By sector, 3.30% of US information-sector postings were AI-related, ahead of professional, scientific and technical services (2.59%), manufacturing (2.02%) and finance and insurance (1.81%).

Four states, and one district

California posted 80,238 AI jobs in 2021, over 2.35 times the 34,021 of Texas in second place, ahead of New York (24,494), Virginia (19,387), Washington (19,253) and Massachusetts (18,430). Raw counts favor big states, so the chapter also ranks AI postings against each state’s own total — and by that measure Washington, D.C. leads, followed by Virginia, Washington and Massachusetts. The capital’s 6,381 AI postings are a fraction of California’s, but they represent the densest AI labor demand in the country.

Skills, not just openings

• Aggregated over 2015–2021, India has the highest relative AI skill penetration at 3.09 times the global average across the same occupations, followed by the United States (2.24), Germany (1.70), China (1.56), Israel (1.52), Canada (1.41) and the United Kingdom (1.40).
• India and the United States lead across all five industries measured — software and IT services, hardware and networking, manufacturing, education and finance. Israel and Canada sit in the top seven of every one of them, and Singapore holds fourth place on the list.
• The education industry shows the widest spread of all: India reaches 3.96 against a global average of 1.00, with the United States at 2.20.
• On gender, among the 15 countries listed the AI skill penetration rate of women is higher than that of men in India, Canada, South Korea, Australia, Finland and Switzerland.

Adoption is a function-level phenomenon

• Across all industries, the functions most likely to have adopted AI in 2021 were service operations (25%), product and/or service development (23%) and marketing and sales (20%). Risk and supply-chain management sat at 13%, manufacturing at 12%, and human resources and strategy and corporate finance at 9% each.
• The heaviest single pockets are narrower than those averages suggest: product and service development inside high tech and telecommunications (45%), service operations in financial services (40%), service operations in high tech and telecommunications (34%) and the risk function in financial services (32%).
• By capability, robotic process automation is the most commonly embedded across all industries (26%), followed by natural language text understanding (24%), computer vision (23%), virtual agents (23%) and deep learning (19%).
• Again, industry concentrates it: natural language text understanding in high tech and telecommunications reaches 34%, robotic process automation reaches 33% in both financial services and automotive and assembly, and natural language text understanding reaches 32% in financial services.

Naming a risk is not the same as acting on it

The chapter asks two questions in sequence — which AI risks does your organization consider relevant, and which are you taking steps to mitigate — and the distance between the answers is the finding. Cybersecurity was named relevant by 55% of respondents and mitigated by 47%. Regulatory compliance: 48% relevant, 36% mitigated, a 12-point gap. Personal and individual privacy: 41% and 28%, a 13-point gap. Explainability: 41% and 27%, a 14-point gap. Equity and fairness: 29% and 19%, a 10-point gap. The risks with the widest gaps are precisely the ones with no established engineering practice behind them.

• Concern about cybersecurity actually fell, from just over 60% of respondents in 2020 to 55% in 2021, while concern about regulatory compliance stayed virtually unchanged.
• Further down the list the absolute numbers get small. Workforce and labor displacement was considered relevant by 26% and mitigated by 15%; national security 14% and 8%; political stability 9% and 4%.
• Organizational reputation — a risk organizations bear directly — sits at 35% relevant and 22% mitigated, still a 13-point gap.

AI is the decade’s dominant specialty

• Artificial intelligence and machine learning accounted for 21.00% of new US CS PhDs in 2020 — the most popular of the survey’s 20 specialties, and the one with the most significant growth from 2010 to 2021 relative to the other 18.
• No other specialty comes close: software engineering is second at 7.30%, then security and information assurance (7.10%), theory and algorithms (7.00%), databases and information retrieval (6.50%) and robotics and vision (6.30%).
• Robotics and vision, the survey’s other explicitly AI-related specialty, gained 1.4 percentage points of share over the past eleven years.
• In absolute numbers, AI/ML PhD graduates grew 72.05% between 2010 and 2020 and robotics and vision graduates 50.91%. Both dipped slightly from 2019 to 2020, which the chapter attributes to the possible impact of the COVID-19 pandemic.

Where the graduates go

• In 2020, 60.24% of new North American AI PhDs went into industry, 24.02% into academia and 1.97% into government. Industry’s share is down from 65.7% in 2019, but the split has been lopsided for a decade.
• Part of the 2020 shift is that more graduates left the continent: the number going abroad on graduation grew from 19 in 2019 to 32 in 2020.
• International students made up 60.5% of new AI PhDs in North America in 2020, down slightly from 64.3% in 2019 — and lower than the 65.1% international share among computing PhDs as a whole.
• Of new international AI PhDs in the United States, 74.20% took jobs in the country and 14.0% took jobs outside it in 2020, up from 8.6% the previous year; the remaining 11.80% is unknown.

Who they are has barely changed

• Women accounted for 20.20% of new AI PhDs and 19.90% of new CS PhDs in North America in 2020 — a share that has moved very little since 2010.
• Among new US-resident AI PhDs from 2010 to 2020, the largest groups on average were white non-Hispanic (65.2%) and Asian (18.8%). Black or African American non-Hispanic graduates averaged around 1.5% and Hispanic graduates 2.9% across those eleven years.
• The 2020 snapshot alone reads 50.86% white non-Hispanic, 30.17% Asian, 6.90% Hispanic and 1.72% Black or African American non-Hispanic.
• Computing PhDs overall show the same pattern: in 2020, 57.50% white non-Hispanic, 24.80% Asian, 4.20% Hispanic and 3.60% Black or African American non-Hispanic. The share of new white PhDs has changed little in eleven years, while Black and Hispanic shares remain significantly lower.